Non-IID and aware federated intrusion detection with PBFT with secured model aggregation for multi institutional healthcare internet of things networks Authors: Sudhakar Sengan, Chin-Shiuh Shieh Published: 2026 ABSTRACT Multi-institutional healthcare Internet of Things (IoT) networks face a core challenge between combined intrusion detection and patient data privacy. Raw traffic records cannot be shared across institutional boundaries, yet local models trained on institution-specific data alone generalize poorly to attack distributions that differ from those practical in real healthcare IoT deployments. Federated Learning (FL) addresses privacy constraints by storing data locally at each institution, but it introduces statistical heterogeneity across institutions. Local data at each institution are non-independent and non-identically distributed due to clinical specialization, protocol diversity, and deployment-scale asymmetry. Standard federated averaging cannot handle this non-IID condition, and detection performance lowers significantly as a result. Existing Federated Intrusion Detection Systems (FIDS) implied that all participating institutions were honest. But this hypothesis cannot hold in real multi-institutional consortia, because Byzantine participants can corrupt the global model by submitting manipulated gradient updates. In this work, a Non-IID-Aware Federated Intrusion Detection System (N-IID-AFIDS) is proposed for multi-institutional healthcare IoT networks and is designed to address both challenges simultaneously. A cluster-weighted aggregation mechanism is used in this N-IID-AFIDS, grouping institutions by distributional similarity through spectral clustering of a Wasserstein-based affinity matrix and applying gradient divergence correction to submitted updates before aggregation. Protocol-aware Deep Sparse Autoencoder (DSAE) adaptation is also part of this model, and it uses local feature normalization based on an institutional protocol mixture and a distribution alignment regularizer. This regularizer operates without raw data exchange across institutions. This work extends practical Byzantine Fault Tolerance (PBFT) consensus from event logging for detection to model update validation. Geometric median-based Byzantine filtering, together with reputation-based participation control, is also added to this model. The proposed model is evaluated on the IoT-Flock and CICIoT2023 datasets across three non-IID severity levels, based on combined skew in quantity, labels, and features. It achieved non-IID detection accuracies of 93.17% on IoT-Flock and 89.84% on CICIoT2023. And this model is the only federated method in the comparison that reports non-IID performance on both datasets simultaneously. It also retains 83.61% accuracy with four Byzantine participants and meets a 16 ms clinical real-time detection constraint, converging in 29–67 rounds, faster than other federated baselines. Introduction The rapid development of Internet of Things (IoT) devices in healthcare has changed how patient monitoring, Clinical Decision-Making System (CDSS), and medical data exchange are carried out. Hospitals and care centers now connect wearable sensors, infusion pumps, imaging systems, and remote monitoring platforms1. All these devices uninterruptedly generate sensitive patient data using heterogeneous protocols. This connectivity has improved care delivery and operational efficiency. But at the same time, it makes the attack region of connected systems much larger2. Attacks such as Distributed Denial-of-Service (DDoS), Man-in-the-Middle (MiTM) interception, and Brute-Force Attacks (BFA) exploit vulnerabilities in connected healthcare systems (HCS)3. It is also found that conventional signature-based Intrusion Detection Systems (IDSs) cannot detect these attacks in real time because they are constantly evolving and growing in scale4. Deep learning (DL)-based IDS has achieved better results than static rule-based methods. This is because they can learn complex temporal and structural attack patterns that are beyond the scope of static rule-based methods5. However, most existing DL-based IDS assume that multiple institutions can share training data centrally. However, this hypothesis does not align with HCS’s actual performance. Regulations such as the Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR) have restricted the sharing of real network traffic across organizational limits in HCS settings6. Therefore, healthcare institutions cannot share traffic logs directly for combined model training because doing so would breach privacy and compliance requirements7. Federated Learning (FL) has addressed data-sharing limitations by enabling institutions to collaboratively train models, with only local updates transmitted for aggregation rather than raw data. Further, FL has been commonly used in HCS. But it is found that its application to healthcare IoT-IDS brings a critical challenge of statistical heterogeneity across institutions8. Network traffic varied significantly across services because large hospitals face a wider range of attack types, unlike a focused health center, which encounters only limited device-specific attacks. This non-IID nature is what has caused client drift in federated training. Each institution’s local updates diverge from the global objective, leading to reduced model accuracy under standard aggregation9. Existing mitigation methods, such as proximal regularization and control variates, had only provided partial solutions. And these methods still cannot handle heterogeneity arising from institutional diversity, protocol differences, and scale changes10. The current federated IDS has not prioritized the security of the aggregation method itself. These IDs operate on the hypothesis that all users are honest. But in multi-institutional settings, this is not realistic because compromised or malicious clients can submit adversarial updates11. When such updates are not properly validated, the global model becomes corrupted, affecting detection reliability. Blockchain (BC)-based systems have enabled secure aggregation by leveraging unchangeable ledgers and Byzantine fault-tolerant consensus12. But existing methods lacked such trust models, or they relied on consensus protocols that introduce latency incompatible with real-time IDS requirements13. In this study, a Non-IID-Aware Federated Intrusion Detection System (N-IID-AFIDS) is proposed to address gaps identified in the existing federated IDS literature. This N-IID-AFIDS is built on cluster-weighted model aggregation, Byzantine-resilient gradient filtering, and PBFT-secured BC validation, serving as its core components for multi-institutional healthcare IoT security. This model adopted a centralized hybrid BC and DL system from earlier work16 and moved it to a fully federated setting. Deep Sparse Autoencoder (DSAE)-based Feature Extraction (FE), a Bidirectional Long Short-Term Memory (BiLSTM) temporal backbone, attack-specific expert modules, and Bayesian Product-of-Experts (PoE) fusion from the earlier network are retained in this model. In addition, federated adaptation is introduced in this research study to address the non-IID challenge and the aggregation security problem simultaneously. This study makes four key contributions, summarized below. The first contribution is the construction of a non-IID data characterization model using Dirichlet-based modeling to represent skew in quantities, labels, and features across healthcare IoT institutions.The second contribution is that the protocol-aware feature normalization stage, together with the Wasserstein-based distribution alignment regularizer, has been integrated into the federated DSAE. And these two components handle embedding conflicts in non-IID environments.The third contribution is a cluster-weighted aggregation mechanism with gradient divergence correction that is also added to this model. It uses spectral clustering, divergence-based weighting, and momentum-based drift adjustment to handle heterogeneous conditions in aggregation.Further, the PBFT-based consensus mechanism was extended in this model to validate model updates through consortium-wide verification. Byzantine-resilient filtering, together with reputation-based participation control, is also added to this mechanism. The first contribution is the construction of a non-IID data characterization model using Dirichlet-based modeling to represent skew in quantities, labels, and features across healthcare IoT institutions. The first contribution is the construction of a non-IID data characterization model using Dirichlet-based modeling to represent skew in quantities, labels, and features across healthcare IoT institutions. The second contribution is that the protocol-aware feature normalization stage, together with the Wasserstein-based distribution alignment regularizer, has been integrated into the federated DSAE. And these two components handle embedding conflicts in non-IID environments. The second contribution is that the protocol-aware feature normalization stage, together with the Wasserstein-based distribution alignment regularizer, has been integrated into the federated DSAE. And these two components handle embedding conflicts in non-IID environments. The third contribution is a cluster-weighted aggregation mechanism with gradient divergence correction that is also added to this model. It uses spectral clustering, divergence-based weighting, and momentum-based drift adjustment to handle heterogeneous conditions in aggregation. The third contribution is a cluster-weighted aggregation mechanism with gradient divergence correction that is also added to this model. It uses spectral clustering, divergence-based weighting, and momentum-based drift adjustment to handle heterogeneous conditions in aggregation. Further, the PBFT-based consensus mechanism was extended in this model to validate model updates through consortium-wide verification. Byzantine-resilient filtering, together with reputation-based participation control, is also added to this mechanism. Further, the PBFT-based consensus mechanism was extended in this model to validate model updates through consortium-wide verification. Byzantine-resilient filtering, together with reputation-based participation control, is also added to this mechanism. This research article is organized as follows: Sect. 2 reviews related work; Sect. 3 defines the system model and problem formulation; Sect. 4 models non-IID heterogeneity; Sect. 5 presents the proposed N-IID-AFIDS; Sect. 6 describes the experimental setup; Sect. 7 analyzes results; and Sect. 8 concludes the study. Literature review Early work on FL for IoT intrusion detection had focused mainly on privacy preservation and secure collaboration between institutions. Begum et al.,17, Govindaram and Jegatheesan31, and Nandanwar and Katarya32 used BC in their work, and the model exchange in these systems is auditable and tamper-resistant across distributed IoT environments. But these methods mainly focus on trust and data privacy. These methods also assume homogeneous data distributions across institutions, failing to account for statistical heterogeneity. Later studies had shifted their attention toward improving model performance and convergence in federated IoT settings. In their study, Abd Elaziz et al.,33 used a tab-transformer in combination with optimization methods for IDS. Albanbay,34 and Khraisat et al.,35 report performance instability under heterogeneous data and limitations of standard aggregation methods. Other works also introduced improvements such as feature selection (FS)36, fog-assisted architectures37, real-time IDS38, and adaptive privacy-preserving methods39. But all of these contributions had focused mainly on efficiency, scalability, and privacy. It is found that none of these works directly handle non-IID data distributions in federated IoT environments. Naz,43 integrated a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) with self-attention for intrusion detection. They achieved 99.93% on the Network Security Laboratory Knowledge Discovery in Databases (NSL-KDD), but this method is centralized, and non-IID federated deployment was not measured. Zhang et al.,42 suggested multidimensional feature fusion with a stacking ensemble across time, space, and load-based traffic features across multiple benchmark datasets, but distributional heterogeneity across institutions is not handled. Li et al.,9 had applied Reinforcement Learning (RL) for energy-efficient anti-jamming routing in UAV networks under adversarial conditions, but a federated IDS setting for healthcare IoT is not addressed in this work. Statistical heterogeneity, or non-IID data, has become a serious problem in FL research. In their study, Ilias et al.,41 found that non-IID conditions lead to worse convergence and detection accuracy in federated IDS. It is found that existing methods have mainly focused on single features, such as label skew, and have not measured the combined effects of quantity and feature, together with protocol-level variability, which is typical in healthcare IoT. Naz,45 invented an Adaptive Federated Learning Framework for Secure and Privacy-Preserving Network Traffic Analysis (AFL-SecNet) that addresses non-IID federated network traffic analysis via adaptive local training and a two-trapdoor cryptosystem for secure aggregation, achieving 97.58% on the University of New South Wales - Network Behavior 2015 (UNSW-NB15) dataset. But multi-institutional healthcare IoT deployment with Byzantine-resilient aggregation and combined quantity, label, and feature skew are not modeled in this model. The research studies are also pointing to the same gaps in the literature. Makris,40 had reported that aggregation robustness, heterogeneity, and scalability are among the key challenges in federated IDs. And Buyuktanir et al., (2024) had also stressed that adaptive aggregation is required for non-IID settings. Hassan et al., (2025) and Pinto et al., (2025) note that BC-based trust integration is nearly what is required for federated IDS. It is found that current models handle privacy, robustness, and trust as separate problems rather than as a unified system. Even after prior developments, a gap that remains in this field is that no existing work has jointly addressed multi-dimensional non-IID heterogeneity arising from institutional, protocol, and scale differences, together with Byzantine-resilient aggregation in adversarial settings. Existing federated IDs either focus solely on improving learning under benign conditions or do not address other critical features. It is found that BC in these models is used only for logging and access control, and the aggregation method is not secured at all. In this work, a unified model is proposed that combines cluster-weighted non-IID-aware aggregation, gradient divergence correction, and Practical Byzantine Fault Tolerance (PBFT)-based model validation for deployment in multi-institutional healthcare IoT environments. System model and problem formulation Multi-institutional healthcare IoT network model Healthcare IoT deployments in this study span multiple autonomous institutions, including hospitals, specialty care centers, diagnostic laboratories, and regional health networks. Each of these institutions operates different IoT devices and operates under strict administrative and privacy limits set by its settings. The network in this model was designed around three layers present across this distributed setting. These layers include the Medical Internet of Things (MIoT) perception layer, institution-level edge-fog processing, and the consortium BC layer used for coordination between institutions. The participating institutions are defined as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:K=\{\mathrm{1,2},\dots\:,K\}$$\end{document}, where \:K=12, and their roles in the experimental setup. Each institution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\in\:K$$\end{document} manages a set of IoT devices, as defined in Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{D}_{k}=\{{d}_{k,1},{d}_{k,2},\dots\:,{d}_{k,\mid\:{D}_{k}\mid\:}\}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{k,j}$$\end{document} is the \:j-th device at institution ‘k’, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mid\:{D}_{k}\mid\:$$\end{document} is the number of devices. Each device generates network traffic samples forming a local dataset, as defined in Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{X}_{k}=\left\{\right({x}_{i},{y}_{i}){\}}_{i=1}^{{N}_{k}}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}\in\:{\mathbb{R}}^{d}$$\end{document} is the feature vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}\in\:\{\mathrm{0,1},\dots\:,C\}$$\end{document} is the class label with 0 signifying benign traffic and \:1 by ‘C’ is attack types, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{k}$$\end{document} is the number of samples at institution ‘k’. The global dataset is expressed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:X={\bigcup\:}_{k=1}^{K}{X}_{k}$$\end{document}, with total sample size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N={\sum\:}_{k=1}^{K}{N}_{k}$$\end{document}. FL model Under FL, each institution \:kmaintains a local model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k}\in\:{\mathbb{R}}^{p}$$\end{document}, while a central server maintains a global model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{g}\in\:{\mathbb{R}}^{p}$$\end{document} of an identical network. Training proceeds in communication rounds where institutions optimize locally on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{k}$$\end{document} and share only model updates. The local objective is defined as in Eq. (3):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{F}_{k}\left(\theta\:\right)=\frac{1}{{N}_{k}}\sum\:_{i=1}^{{N}_{k}}l\left(f\right({x}_{i};\theta\:),{y}_{i})&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:l(\cdot\:,\cdot\:)$$\end{document} is the cross-entropy loss, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f({x}_{i};\theta\:)$$\end{document} is the model prediction, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{k}$$\end{document} is the local dataset size as defined in Eq. (2). The global objective is specified as in Eq. (4):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:F\left(\theta\:\right)=\sum\:_{k=1}^{K}\frac{{N}_{k}}{N}{F}_{k}\left(\theta\:\right)&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{N}_{k}}{N}$$\end{document} is the data-proportional contribution of institution ‘k’. Under standard federated averaging, the global model update at round ‘t’ follows Eq. (5):5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\theta\:}_{t+1}^{g}=\sum\:_{k=1}^{K}\frac{{N}_{k}}{N}{\theta\:}_{k,t}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k,t}$$\end{document} is the local model after ‘τ’ gradient steps. Non-IID data distribution formulation The main challenge in this setting is statistical heterogeneity across institutions. It is because local datasets at each institution follow their own distributions, rather than a common IID distribution. This work measures the three forms of non-IID heterogeneity. Quantity skew captures imbalance in dataset sizes across institutions, measured by the coefficient of variation as defined in Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\mathrm{C}\mathrm{V}}_{N}=\frac{{\sigma\:}_{N}}{{\mu\:}_{N}}&\:&\:\end{array}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{N}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{N}$$\end{document} is the standard deviation and mean of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{N}_{1},{N}_{2},\dots\:,{N}_{K}\right\}$$\end{document}. Higher values indicate greater imbalance.Label Skew is a variation in class distributions, where the class probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{q}_{k,c}={P}_{k}(y=c)$$\end{document} differs across institutions.This is modelled using a Dirichlet distribution as defined in Eq. (7):7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\left({q}_{k,0},{q}_{k,1},\dots\:,{q}_{k,C})\sim\:\mathrm{D}\mathrm{i}\mathrm{r}(\alpha\:\cdot\:p\right)&\:&\:\end{array}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p=({p}_{0},{p}_{1},\dots\:,{p}_{C})$$\end{document} is the global class distribution, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:>0$$\end{document} controls heterogeneity severity. Lower ‘α’ induces stronger skew.Feature Skew captures differences in input distributions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}\left(x\right)$$\end{document} due to protocol variability.It is quantified using the Wasserstein distance as defined in Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{1} \left( {P_{k} ,P_{{k^{\prime } }} } \right) = \mathop {{\mathrm{inf}}}\limits_{{\gamma \in \Pi \left( {P_{k} ,P_{{k^{\prime } }} } \right)}} {\mathbb{E}}_{{\left( {x,x^{\prime } } \right)\sim \gamma }} \left[ {\left\| {x - x^{\prime } } \right\|_{1} } \right]$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Pi\:}({P}_{k},{P}_{{k}^{{\prime\:}}})$$\end{document} is the set of joint distributions with marginals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{{k}^{{\prime\:}}}$$\end{document}. Quantity skew captures imbalance in dataset sizes across institutions, measured by the coefficient of variation as defined in Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\mathrm{C}\mathrm{V}}_{N}=\frac{{\sigma\:}_{N}}{{\mu\:}_{N}}&\:&\:\end{array}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{N}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{N}$$\end{document} is the standard deviation and mean of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{N}_{1},{N}_{2},\dots\:,{N}_{K}\right\}$$\end{document}. Higher values indicate greater imbalance. Quantity skew captures imbalance in dataset sizes across institutions, measured by the coefficient of variation as defined in Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\mathrm{C}\mathrm{V}}_{N}=\frac{{\sigma\:}_{N}}{{\mu\:}_{N}}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{N}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{N}$$\end{document} is the standard deviation and mean of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{N}_{1},{N}_{2},\dots\:,{N}_{K}\right\}$$\end{document}. Higher values indicate greater imbalance. Label Skew is a variation in class distributions, where the class probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{q}_{k,c}={P}_{k}(y=c)$$\end{document} differs across institutions. Label Skew is a variation in class distributions, where the class probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{q}_{k,c}={P}_{k}(y=c)$$\end{document} differs across institutions. This is modelled using a Dirichlet distribution as defined in Eq. (7):7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\left({q}_{k,0},{q}_{k,1},\dots\:,{q}_{k,C})\sim\:\mathrm{D}\mathrm{i}\mathrm{r}(\alpha\:\cdot\:p\right)&\:&\:\end{array}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p=({p}_{0},{p}_{1},\dots\:,{p}_{C})$$\end{document} is the global class distribution, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:>0$$\end{document} controls heterogeneity severity. Lower ‘α’ induces stronger skew. This is modelled using a Dirichlet distribution as defined in Eq. (7):7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\left({q}_{k,0},{q}_{k,1},\dots\:,{q}_{k,C})\sim\:\mathrm{D}\mathrm{i}\mathrm{r}(\alpha\:\cdot\:p\right)&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p=({p}_{0},{p}_{1},\dots\:,{p}_{C})$$\end{document} is the global class distribution, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:>0$$\end{document} controls heterogeneity severity. Lower ‘α’ induces stronger skew. Feature Skew captures differences in input distributions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}\left(x\right)$$\end{document} due to protocol variability.It is quantified using the Wasserstein distance as defined in Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{1} \left( {P_{k} ,P_{{k^{\prime } }} } \right) = \mathop {{\mathrm{inf}}}\limits_{{\gamma \in \Pi \left( {P_{k} ,P_{{k^{\prime } }} } \right)}} {\mathbb{E}}_{{\left( {x,x^{\prime } } \right)\sim \gamma }} \left[ {\left\| {x - x^{\prime } } \right\|_{1} } \right]$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Pi\:}({P}_{k},{P}_{{k}^{{\prime\:}}})$$\end{document} is the set of joint distributions with marginals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{{k}^{{\prime\:}}}$$\end{document}. Feature Skew captures differences in input distributions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}\left(x\right)$$\end{document} due to protocol variability. It is quantified using the Wasserstein distance as defined in Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_{1} \left( {P_{k} ,P_{{k^{\prime } }} } \right) = \mathop {{\mathrm{inf}}}\limits_{{\gamma \in \Pi \left( {P_{k} ,P_{{k^{\prime } }} } \right)}} {\mathbb{E}}_{{\left( {x,x^{\prime } } \right)\sim \gamma }} \left[ {\left\| {x - x^{\prime } } \right\|_{1} } \right]$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\Pi\:}({P}_{k},{P}_{{k}^{{\prime\:}}})$$\end{document} is the set of joint distributions with marginals \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{{k}^{{\prime\:}}}$$\end{document}. Threat and attack taxonomy The threat model in this work considered two types of adversaries: external attackers targeting the healthcare IoT setup and Byzantine participants aiming to disrupt it. External adversaries operate on three levels. Network-level attacks involve Distributed Denial-of-Service (DDoS) floods that disrupt communication between MIoT devices and fog nodes. Session-level attacks are carried out through Man-in-the-Middle (MiTM) interception, exploiting weaknesses in encryption or certificate validation. Authentication-level attacks exploit BFA to target device interfaces. All these attack types are encoded in the label space \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}\in\:\left\{\mathrm{0,1},\mathrm{2,3}\right\}$$\end{document} and 0 is for benign traffic, while 1–3 are attack types. Byzantine anomalies in this model are treated as a subset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:B\subset\:K$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mid\:B\mid\:\le\:f0$$\end{document} is the quantity concentration parameter, and ‘N’ is the total sample count. Lower ‘β’ generates more unequal size distributions where a small number of institutions hold the majority of training samples. The resulting institution sizes satisfy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sum\:}_{k=1}^{K}{N}_{k}=N$$\end{document} by construction. The degree of quantitative imbalance is measured using the coefficient of variation from Eq. (6). The Gini coefficient provides a complementary bounded, scale-invariant measure of quantitative imbalance, as assumed in Eq. (14):14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{N}=\frac{{\sum\:}_{k=1}^{K}{\sum\:}_{{k}^{{\prime\:}}=1}^{K}\mid\:{N}_{k}-{N}_{{k}^{{\prime\:}}}\mid\:}{2K{\sum\:}_{k=1}^{K}{N}_{k}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{N}\in\:\left[\mathrm{0,1}\right]$$\end{document} with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{N}=0$$\end{document} signifying perfect size equality and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{N}\to\:1$$\end{document} signifying extreme concentration of data in a single institution. The impact of quantity skew on aggregation is isolated experimentally by fixing the label distribution to the global marginal ‘p’ and varying only the variable ′β′, enabling precise attribution of performance degradation to size asymmetry. Label skew modelling across institutional node types Label skew is the most consequential form of non-IID heterogeneity for IDS because it directly governs which attack classes each institution’s local model learns to recognize. When a local dataset contains few or no samples of a particular attack class, the corresponding local model develops poor sensitivity to that class, and naive aggregation propagates this blindness to the global model. The Dirichlet-based label distribution model from Eq. (7) is applied per institution to assign class-conditional sampling proportions. Institutions are partitioned into functional groups reflecting distinct clinical roles within a healthcare network. Each group is characterized by a structurally motivated prior over attack classes, reflecting which threat types are most prevalent in that institutional context. The group-level prior vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{p}}^{\left(g\right)}$$\end{document} for group ‘\:g’ is defined as in Eq. (15):15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{p}}^{\left(g\right)}=\mathrm{Soft}\mathrm{M}\mathrm{ax}({\mathbf{v}}^{\left(g\right)}/{T}_{prior})$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{v}}^{\left(g\right)}\in\:{\mathbb{R}}^{C+1}$$\end{document} is a group-specific preference vector encoding which attack classes are most prevalent in that institutional context, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{prior}>0$$\end{document} is a temperature parameter controlling the sharpness of the group prior. The final local class proportions for institution ‘\:k’ in group ‘\:g’ are attained by sampling from Eq. (7) using \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{p}}^{\left(g\right)}$$\end{document} as the base vector. The divergence between local and global class distributions is quantified using the Jensen-Shannon divergence specified as in Eq. (16):16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{JSD}({P}_{k}\parallel\:P)=\frac{1}{2}\mathrm{KL}({P}_{k}\parallel\:M)+\frac{1}{2}\mathrm{KL}(P\parallel\:M)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:M=\frac{1}{2}({P}_{k}+P)$$\end{document} is the mixture distribution, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{KL}(\cdot\:\parallel\:\cdot\:)$$\end{document} is the Kullback-Leibler divergence, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}$$\end{document} is the local label distribution of institution ‘k’, and ‘p’ is the global label distribution. The mean \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{JSD}$$\end{document} across all institutions, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overline{{{\mathrm{JSD}}}} = \frac{1}{K}\sum\limits_{{k = 1}}^{K} {{\mathrm{JSD}}(P_{k} \parallel \:P)}$$\end{document}, serves as the scalar heterogeneity measure. Feature skew modelling across protocol environments Feature skew arises from protocol-level differences in traffic features and is modeled by conditioning the input feature distribution on each institution’s dominant protocol profile. Each institution ‘k’ is assigned a protocol mixture weight vector as in Eq. (17):17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\pi\:}_{k}=({\pi\:}_{k,\:\mathrm{MQTT}},{\pi\:}_{k,\:\mathrm{CoAP}},{\pi\:}_{k,\:\mathrm{HL7}},{\pi\:}_{k,\:\mathrm{DICOM}})$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\pi\:}_{k,j}\ge\:0\:$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sum\:}_{j}{\pi\:}_{k,j}=1$$\end{document}, with each component representing the fraction of traffic at institution ‘k’ generated by the corresponding protocol. These weights rule the sampling mixture from which the local feature distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}\left(x\right)$$\end{document} is made. The feature covariance shift between two institutions is measured using the Frobenius norm of the difference between their sample covariance matrices, as stated in Eq. (18):18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{{\Sigma\:}}(k,{k}^{{\prime\:}})=\parallel\:{{\Sigma\:}}_{k}-{{\Sigma\:}}_{{k}^{{\prime\:}}}{\parallel\:}_{F}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Sigma\:}}_{k}=\frac{1}{{N}_{k}}{\sum\:}_{i=1}^{{N}_{k}}({x}_{i}-{\mu\:}_{k}\left)\right({x}_{i}-{\mu\:}_{k}{)}^{T}$$\end{document} is the trial covariance of the local dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{X}}_{k}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{k}$$\end{document} is its experimental mean, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\parallel\:\cdot\:{\parallel\:}_{F}$$\end{document} is the Frobenius norm. Proposed federated IDS Overall workflow of proposed model The proposed model had combined FL and BC-secured aggregation within a layered network, keeping local detection and global coordination separate. This model extends the prior hybrid BC + DL model16 into a multi-institutional setting. The local pipeline at each institution runs independently and is built with a DSAE-based FE, a BiLSTM-based temporal backbone with expert modules, and Bayesian fusion. Two new features introduced in this model are a non-IID-aware aggregation mechanism and a PBFT-secured model update protocol. Understanding of inter-layer data flow is also important in this model. And this is because all components operate within a tightly coupled pipeline—the model’s overall workflow (Fig. 1). Fig. 1Overall workflow of proposed N-IID-AFIDS. Overall workflow of proposed N-IID-AFIDS. The research of this N-IID-AFIDS had 3 tiers as described in Sect. 3.1. At the perception and edge–fog tier, each institution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\in\:K$$\end{document} is processing local traffic \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{k}$$\end{document} from Eq. (2) through the pipeline that is dedicated to that institution. This pipeline includes traffic preprocessing, DSAE-based FE, parallel BiLSTM temporal modeling with attack-specific expert modules, and Bayesian PoE fusion, and provides calibrated detection results. It runs continuously at each institution. And the real-time decisions generated by this pipeline are independent of global aggregation. At the federation coordination tier, institutions send local model parameters to the central server at each communication round ‘t’. The central server then applies non-IID-aware cluster-weighted aggregation and gradient divergence that comes from heterogeneity in Eq. (7) and Eq. (8), which is corrected by this method. The updated global model is then sent back to institutions for the next round. The process follows a three-phase cycle. In the local training phase, each institution starts from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t}^{g}$$\end{document} and performs ‘τ’ gradient steps on its local objective (Eq. (19)), generating:19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\theta\:}_{k,t}={\theta\:}_{t}^{g}-\eta\:\sum\:_{{\tau\:}^{{\prime\:}}=1}^{\tau\:}\nabla\:{F}_{k}\left({\theta\:}_{k,t,{\tau\:}^{{\prime\:}}-1}\right)&\:&\:\end{array}$$\end{document} where ‘η’ is the learning rate and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k,t,0}={\theta\:}_{t}^{g}$$\end{document}. In the update submission phase, the model update is computed as Eq. (20)20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{{\Delta\:}}_{k,t}={\theta\:}_{k,t}-{\theta\:}_{t}^{g}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}\in\:{\mathbb{R}}^{p}$$\end{document} is transmitted to the server, ensuring privacy by avoiding raw data sharing as constrained in Eq. (12). In the global aggregation phase, the server combines updates using the proposed non-IID-aware mechanism to produce \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document}, which is redistributed to all institutions. At the BC tier, model updates from Eq. (20) were wrapped as PBFT-validated transactions before aggregation. Only updates accepted by the honest majority are incorporated into the aggregation method. Smart Contracts (SC) enforce input rules, penalize anomalous updates, and maintain an immutable audit trail. In each round, institutions receive \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t}^{g}$$\end{document}, do local training as per Eq. (19), compute updates as per Eq. (20), and submit them as blockchain (BC) transactions. PBFT consensus then checks and confirms these updates. After this aggregation server applies a non-IID-aware mechanism from Sect. 5.4 and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document} is formed as output. The updated model is sent back to all institutions. This method repeats until the condition in Eq. (4) is met or the maximum number of rounds is reached. The local detection pipeline operates continuously throughout all rounds. Non-IID-adapted DSAE The first stage of the local pipeline performs FE, transforming heterogeneous IoT traffic into compact latent representations. A DSAE (Fig. 2) is employed, improved with two federated adaptations: protocol-aware local normalization to account for institution-specific protocol mixtures and a distribution alignment regularizer to promote cross-institutional embedding consistency without sharing raw data. Fig. 2Non-IID-adapted DSAE. Non-IID-adapted DSAE. DSAE and sparsity regularization The DSAE employs a hierarchical encoder-decoder structure of ‘L’ encoding layers. The encoding operation at layer ‘l’, Eq. (21):21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:h_{k}^{{\left( l \right)}} = f\left( {W_{k}^{{\left( l \right)}} h_{k}^{{\left( {l - 1} \right)}} + b_{k}^{{\left( l \right)}} } \right),l = 1,2, \ldots \:,L$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{k}^{\left(0\right)}={\stackrel{\sim}{x}}_{i}$$\end{document} is the pre-processed input at institution \:k, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{k}^{\left(l\right)}\in\:{\mathbb{R}}^{{d}_{l}\times\:{d}_{l-1}}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{k}^{\left(l\right)}\in\:{\mathbb{R}}^{{d}_{l}}$$\end{document} are the weight matrix and bias vector at layer \:l, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f(\cdot\:)$$\end{document} is the ReLU activation function, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{l}$$\end{document} is the number of hidden units in layer ‘l’ with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{L}\ll\:{d}_{0}$$\end{document} enforcing dimensionality reduction. The symmetric decoder mirrors the encoder to reconstruct the input from the bottleneck illustration \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k,i}={h}_{k}^{\left(L\right)}\in\:{\mathbb{R}}^{{d}_{L}}$$\end{document}. The reconstruction loss across \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{k}$$\end{document} local samples are expressed as in Eq. (22):22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{rec,k}=\frac{1}{{N}_{k}}{\sum\:}_{i=1}^{{N}_{k}}{\parallel\:{\stackrel{\sim}{x}}_{i}-{\widehat{x}}_{i}\parallel\:}^{2}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{x}}_{i}$$\end{document} is the decoder output for input \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{x}}_{i}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\parallel\:\cdot\:\parallel\:$$\end{document} is the Euclidean norm. To prevent trivial identity mapping and help discriminative feature selectivity, a sparsity drawback is imposed on the mean activation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{\rho\:}}_{k,j}^{\left(l\right)}$$\end{document} of hidden unit ‘j’ at layer ‘l’ over the local dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{X}}_{k}$$\end{document}. The layer-wise sparsity drawback is defined as in Eq. (23):23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Omega\:}}_{k}^{\left(l\right)}={\sum\:}_{j=1}^{{d}_{l}}\mathrm{KL}\left(\rho\:\parallel\:{\widehat{\rho\:}}_{k,j}^{\left(l\right)}\right)$$\end{document} where ‘ρ’ is the target sparsity level and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{KL}(\rho\:\parallel\:\widehat{\rho\:})=\rho\:\mathrm{L}\mathrm{o}\mathrm{g}\frac{\rho\:}{\widehat{\rho\:}}+(1-\rho\:)\mathrm{L}\mathrm{o}\mathrm{g}\frac{1-\rho\:}{1-\widehat{\rho\:}}$$\end{document} is the KL divergence between the target and experimental activation. The standard DSAE objective aggregates Eq. (22) and Eq. (23) across all encoding layers as stated in Eq. (24):24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{DSAE,k}^{std}={L}_{rec,k}+\beta\:{\sum\:}_{l=1}^{L}{{\Omega\:}}_{k}^{\left(l\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:>0$$\end{document} is the sparsity regularization coefficient. Protocol-aware local feature normalization In the centralized setting, all traffic is drawn from a single dataset and uniformly normalized before DSAE encoding. Under federation, this hypothesis fails due to protocol-level differences in feature scales and distributions across institutions, rendering global normalization ineffective and hypothetically distorting diagnostic data. The proposed protocol-aware normalization instead conditions input statistics on the local protocol mixture \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\pi\:}_{k}$$\end{document} from Eq. (17) before encoding in Eq. (21). For each input dimension ‘j’, the normalization transform at institution ‘k’ is defined as in Eq. (25):25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{x}}_{i,j}=\frac{{x}_{i,j}-{\mu\:}_{k,j}^{\pi\:}}{{\sigma\:}_{k,j}^{\pi\:}+\varepsilon}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{k,j}^{\pi\:}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{k,j}^{\pi\:}\:$$\end{document}are the protocol-weighted mean and standard deviation of feature ‘j’ at institution \:k, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varepsilon>0$$\end{document} is a numerical stability constant. The protocol-weighted mean is computed as in Eq. (26):26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{k,j}^{\pi\:}={\sum\:}_{r}{\pi\:}_{k,r}\cdot\:{\mu\:}_{r,j}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\pi\:}_{k,r}$$\end{document} is the weight of protocol ‘r’ in the local mixture from Eq. (17), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{r,j}$$\end{document} is the empirical mean of feature ‘j’ computed over traffic samples of protocol ‘r’ within \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{X}}_{k}$$\end{document}. The standard deviation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{k,j}^{\pi\:}$$\end{document} is computed analogously by substituting the change under each protocol. Distribution alignment regularizer The distribution alignment regularizer presented here penalizes the divergence between the local embedding distribution at an institution ‘k’ and a global reference embedding distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}^{g}\left(z\right)$$\end{document} predictable from the prior round’s global model. The regularizer is defined as in Eq. (27):27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Omega\:}}_{align,k}={W}_{1}\left({P}_{k}\left(z\right),{P}^{g}\left(z\right)\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{1}(\cdot\:,\cdot\:)$$\end{document} is the first Wasserstein distance from Eq. (8), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{k}\left(z\right)$$\end{document} is the experimental distribution of bottleneck embeddings formed by the local DSAE at institution ‘k’, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}^{g}\left(z\right)$$\end{document}is the reference embedding distribution computed from the global model applied to a small shared correction set held at the aggregation server. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{1}$$\end{document} in Eq. (27), is differentiable with respect to the DSAE limits using its dual formulation, and it is incorporated directly into the training objective. The complete Non-IID-Adapted DSAE local objective is defined as in Eq. (28):28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{DSAE,k}={L}_{DSAE,k}^{std}+{\lambda\:}_{align}\cdot\:{{\Omega\:}}_{align,k}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{DSAE,k}^{std}$$\end{document} is the standard objective from Eq. (24), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Omega\:}}_{align,k}$$\end{document} is the alignment drawback from Eq. (27), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{align}>0$$\end{document} is a regularization coefficient governing the trade-off between local reform fidelity and cross-institutional embedding reliability. Local training procedure The Non-IID-adapted DSAE at the institution ‘k’ is trained in 2 phases. In the pre-training phase, each encoding layer is optimized autonomously using the single-layer form of Eq. (24) without alignment. In the fine-tuning phase, the full objective in Eq. (28) is jointly minimized across all layers, while the global reference distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}^{g}\left(z\right)$$\end{document} from Eq. (27) is updated each round using \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t}^{g}$$\end{document} Eq. (19). The final latent embedding for each input is attained as in Eq. (29):29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{z}_{k,i}={h}_{k}^{\left(L\right)}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{z}_{k,i}\in\:{\mathbb{R}}^{{d}_{L}}$$\end{document} is the compact model passed to the detection module. The transformed dataset is defined in Eq. (30):30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{Z}_{k}=\left\{\right({z}_{k,i},{y}_{i}){\}}_{i=1}^{{N}_{k}}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{k}$$\end{document} is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{X}_{k}$$\end{document} from Eq. (2), attaining dimensionality reduction (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{L}\ll\:d$$\end{document}) while preserving a discriminative network across heterogeneous protocols. Local detection engine The latent embeddings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{k}$$\end{document} from Eq. (30) are input to the local detection engine, which generates standardized attack classifications at each institution ‘k’. The engine comprises four parallel components: a BiLSTM temporal backbone and three focused expert modules for DDoS, MiTM, and Brute-Force Attacks (BFA), with their outputs integrated via a Bayesian PoE fusion layer. BiLSTM temporal backbone The BiLSTM backbone method’s sequence of embeddings \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{z}_{k,\tau\:}{\}}_{\tau\:=t-T+1}^{t}\right.$$\end{document} Eq. (29) captures bidirectional temporal dependencies. Forward and backward passes generate hidden states that are concatenated at each timestep, as in Eq. (31):31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{h}_{\tau\:}=[{h}_{\tau\:}^{\left(f\right)};{h}_{\tau\:}^{\left(b\right)}]&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{\tau\:}\in\:{\mathbb{R}}^{2{d}_{h}}.$$\end{document}. An attention mechanism aggregates these illustrations into a sequence summary \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{\prime }{h}$$\end{document}, which is used to compute class probabilities as defined in Eq. (32):32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{p}_{\mathrm{backbone}}=\mathrm{Soft}\mathrm{M}\mathrm{ax}({W}_{c}\stackrel{\prime }{h}+{b}_{c})&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{backbone}}\in\:{\mathbb{R}}^{C+1}$$\end{document}, these are the probabilities for the attack and benign classes. Attack-specific expert modules Three expert modules operate on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Z}_{k}$$\end{document} from Eq. (30), each targeting a precise attack type. For DDoS detection, RF-based FS retains a subset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{\mathrm{ddos}}$$\end{document}, followed by an MLP classifier producing:33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{p}_{\mathrm{ddos}}=\mathrm{sigmoid}\left({W}_{\mathrm{out}}^{\left(\mathrm{ddos}\right)}{h}_{{L}_{\mathrm{ddos}}}^{\left(\mathrm{ddos}\right)}+{b}_{\mathrm{out}}^{\left(\mathrm{ddos}\right)}\right)&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{ddos}}\in\:\left[\mathrm{0,1}\right]$$\end{document} is the DDoS probability, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{{L}_{\mathrm{ddos}}}^{\left(\mathrm{ddos}\right)}$$\end{document} is the final hidden layer network, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{\mathrm{out}}^{\left(\mathrm{ddos}\right)},{b}_{\mathrm{out}}^{\left(\mathrm{ddos}\right)}$$\end{document}are output limits. For MiTM detection, the attacker’s interception behavior unfolds across multiple data packets sequentially and adaptively, and this temporal adversarial nature makes Markov Decision Process (MDP) modeling more suitable than static classification for this attack type. For this, a DQN-based agent models detection as an MDP, and the probability is derived as:34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{p}_{\mathrm{mitm}}=\underset{a\in\:\left\{{a}_{\mathrm{MiTM}}\right\}}{\mathrm{m}\mathrm{a}\mathrm{x}}Q(s,a;\theta\:)&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{mitm}}\in\:\left[\mathrm{0,1}\right]$$\end{document} is the MiTM probability and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Q(s,a;\theta\:)$$\end{document} is the learned action-value function. For BFA, repeated credential attempts against authentication interfaces generate fixed-length, high-frequency burst patterns in the authentication sequence matrix, and these local spatial regularities are what Conv2D layers are more effective at capturing than temporal models. Authentication sequences are processed using Conv2D and pooling layers, generating:35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{p}_{\mathrm{brute}}=\mathrm{sigmoid}({W}_{fc}{F}_{\mathrm{flat}}+{b}_{fc})&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{\mathrm{flat}}$$\end{document} is the flattened FS, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{brute}}\in\:\left[\mathrm{0,1}\right]$$\end{document} is the BFA probability, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{fc},{b}_{fc}$$\end{document} are fully connected layer parameters. Bayesian product-of-experts fusion The outputs of the 4 modules—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{backbone}}$$\end{document} from Eq. (32), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{ddos}}$$\end{document} from Eq. (33), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{mitm}}$$\end{document} from Eq. (34), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{brute}}$$\end{document} from Eq. (35) are fused into a standardized multi-class distribution. Binary expert outputs are first expanded to full class distributions, assigning ‘0’ probability outside their possibility. Fusion is performed using a weighted PoE formulation as defined in Eq. (36):36\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{p}_{\mathrm{fused}}^{\left(c\right)}\propto\:\prod\:_{k=1}^{4}{\left({p}_{k}^{\left(c\right)}\right)}^{{w}_{k}^{\left(c\right)}}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{k}^{\left(c\right)}$$\end{document} does PoE assign the probability ‘k’to class ‘c’, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{k}^{\left(c\right)}$$\end{document} are learned reliability weights. The fused distribution is normalized and standardized using temperature scaling, followed by context-aware modulation as defined in Eq. (37):37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{p}_{\mathrm{final},k}^{\left(c\right)}={p}_{\mathrm{cal}}^{\left(c\right)}\cdot\:{\phi\:}_{\mathrm{context}}(\mathrm{device\_type},\mathrm{clinical\_status})&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{\mathrm{cal}}^{\left(c\right)}$$\end{document} is the calibrated probability and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\phi\:}_{\mathrm{context}}$$\end{document} is the modulation function. Predictive uncertainty is quantified using entropy as defined in Eq. (38):38\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:H\left({p}_{\mathrm{final},k}\right)=-\sum\:_{c=0}^{C}{p}_{\mathrm{final},k}^{\left(c\right)}\mathrm{L}\mathrm{o}\mathrm{g}{p}_{\mathrm{final},k}^{\left(c\right)}&\:&\:\end{array}$$\end{document} where higher entropy specifies uncertain predictions requiring review. The final output is the tuple \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({p}_{\mathrm{final},k},{\mathrm{decision}}_{k},H\left({p}_{\mathrm{final},k}\right)\right)$$\end{document}, which feeds the aggregation mechanism (Sect. 5.4) and the BC logging layer (Sect. 5.5). Non-IID-aware global aggregation mechanism Limitations of standard FedAvg under non-IID conditions Client drift under non-IID data arises when local updates follow institution-specific loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{k}\left(\theta\:\right)$$\end{document} instead of the global objective \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:F\left(\theta\:\right)$$\end{document} from Eq. (4), with divergence increasing with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{JSD}({P}_{k}\parallel\:P)$$\end{document}, Eq. (16). The predictable gradient divergence is limited as in Eq. (39):39\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:\mathbb{E}\left[\parallel\:\nabla\:{F}_{k}\left(\theta\:\right)-\nabla\:F\left(\theta\:\right){\parallel\:}^{2}\right]\le\:{G}^{2}\cdot\:\mathrm{JSD}({P}_{k}\parallel\:P)&\:&\:\end{array}$$\end{document} where ‘G’ is a gradient bound constant. Under FedAvg (Eq. (5)), this divergence collects across institutions, generating the convergence bound in Eq. (40):40\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${{\mathbb{E}}\left[ {\left\| {\theta _{{t + 1}}^{g} - \theta ^{*} } \right\|^{2} } \right] \le \sum\limits_{{k = 1}}^{K} {\frac{{N_{k} }}{N}{\mathrm{G}}^{{\mathrm{2}}} \cdot {\mathrm{JSD}}({\mathrm{P}}_{{\mathrm{k}}} \parallel {\mathrm{P}}) \cdot \tau ^{{\mathrm{2}}} \eta ^{{\mathrm{2}}} } }$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{*}$$\end{document}minimizes Eq. (4), ‘η’ is the learning rate (Eq. (19)), and ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}$$\end{document}’ is the number of local steps. This proof of error grows with heterogeneity and quadratically with ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau$$\end{document}’ and ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta$$\end{document}’, leading to severe degradation at high skew (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.1$$\end{document}). Cluster-weighted aggregation method The cluster-weighted aggregation replaces the data-proportional weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{N}_{k}}{N}$$\end{document} in Eq. (5) with distributional similarity weights to reduce the influence of highly skewed institutions. This mechanism is operating in four sequential steps. First, pairwise distributional affinity between institutions is computed using Jensen-Shannon divergence to form the affinity matrix.Second, spectral clustering is applied to the affinity matrix to partition institutions into distributional-similarity groups.Third, within-cluster updates are aggregated using data-proportional weights restricted to honest institutions.Fourth, inter-cluster weights are computed based on each cluster’s mean divergence from the global distribution, and these weights are used to combine cluster-level updates into the global model. First, pairwise distributional affinity between institutions is computed using Jensen-Shannon divergence to form the affinity matrix. First, pairwise distributional affinity between institutions is computed using Jensen-Shannon divergence to form the affinity matrix. Second, spectral clustering is applied to the affinity matrix to partition institutions into distributional-similarity groups. Second, spectral clustering is applied to the affinity matrix to partition institutions into distributional-similarity groups. Third, within-cluster updates are aggregated using data-proportional weights restricted to honest institutions. Third, within-cluster updates are aggregated using data-proportional weights restricted to honest institutions. Fourth, inter-cluster weights are computed based on each cluster’s mean divergence from the global distribution, and these weights are used to combine cluster-level updates into the global model. Fourth, inter-cluster weights are computed based on each cluster’s mean divergence from the global distribution, and these weights are used to combine cluster-level updates into the global model. Institutions are grouped using pairwise Jensen–Shannon divergence from Eq. (16), with the affinity matrix defined as in Eq. (41):41\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${A_{{k,k^{\prime } }} = {\mathrm{Exp}}\left( { - \frac{{{\mathrm{JSD}}(P_{k} \parallel P_{{k^{\prime } }} )}}{{\sigma _{{{\mathrm{aff}}}}^{2} }}} \right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left.{A}_{k,{k}^{{\prime\:}}}\in\:(\mathrm{0,1}\right]$$\end{document} is the affinity between institutions \:kand \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{k}^{{\prime\:}}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{\mathrm{aff}}^{2}>0$$\end{document} is a bandwidth parameter controlling sensitivity to distributional distance. Spectral clustering is then applied using the normalized Laplacian defined as in Eq. (42):42\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:L=I-{D}^{-\frac{1}{2}}A{D}^{-\frac{1}{2}}&\:&\:\end{array}$$\end{document} where ‘D’ is the diagonal degree matrix with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}_{kk}={\sum\:}_{{k}^{{\prime\:}}}{A}_{k,{k}^{{\prime\:}}}$$\end{document} and ‘l’ is the identity matrix; the resulting spectral embedding partitions institutions into clusters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{C}_{g}{\}}_{g=1}^{G}\right.$$\end{document}. Within each cluster, updates are aggregated as in Eq. (43):43\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{{\Delta\:}}_{g}^{\mathrm{cluster}}=\frac{\sum\:_{k\in\:{C}_{g}}{N}_{k}\cdot\:{{\Delta\:}}_{k,t}}{\sum\:_{k\in\:{C}_{g}}{N}_{k}}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}$$\end{document} is the local update from Eq. (20) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{k}$$\end{document} is the dataset size at institution ‘k’. The cluster-level updates are then combined into the global model as in Eq. (44):44\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{\theta\:}_{t+1}^{g}={\theta\:}_{t}^{g}+\sum\:_{g=1}^{G}{\omega\:}_{g}\cdot\:{{\Delta\:}}_{g}^{\mathrm{cluster}}&\:&\:\end{array}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t}^{g}$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document} are the global parameters before and after aggregation. The inter-cluster weights are defined as in Eq. (45):45\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\omega _{g} = \frac{{{\mathrm{exp}}\left( { - \overline{{{\mathrm{JSD}}}} _{g} /T_{{{\mathrm{clust}}}} } \right)}}{{\sum\nolimits_{{g^{\prime } = 1}}^{G} {{\mathrm{exp}}\left( {\: - \overline{{{\mathrm{JSD}}}} _{{g^{\prime } }} /T_{{{\mathrm{clust}}}} } \right)} }}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overline{{{\mathrm{JSD}}}} _{g} = \frac{1}{{{\mid }C_{g} {\mid }}}\sum\limits_{{k \in C_{g} }} {{\mathrm{JSD}}(P_{k} \parallel P)}$$\end{document} is the mean divergence of cluster ‘g’ from the global distribution, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{\mathrm{clust}}>0$$\end{document} is a temperature parameter controlling weight sharpness. Gradient divergence correction The gradient divergence correction term presented here modifies each institution’s submitted update \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}\:$$\end{document}by subtracting a predictable drift component before the cluster aggregation in Eq. (43) is applied. This correction is applied after cluster grouping from Sect. 5.4.2 is completed and before the corrected updates are passed into Eq. (43), so that drift-corrupted updates from skewed institutions are removed before within-cluster aggregation rather than after it. The drift estimate at the institution ‘k’ is computed as the divergence between the local update direction and the cluster centroid direction from the previous round as specified in Eq. (46):46\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\delta\:}_{k,t}={{\Delta\:}}_{k,t}-{{\Delta\:}}_{k,t-1}^{corrected}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t-1}^{corrected}$$\end{document} is the drift-corrected update from institution ‘k’ in the previous round, maintained at the aggregation server as a running estimate. The corrected update for institution ‘k’ in round ‘t’ is defined as in Eq. (47):47\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}^{corrected}={{\Delta\:}}_{k,t}-{\mu\:}_{corr}\cdot\:{\delta\:}_{k,t}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{corr}\in\:\left[\mathrm{0,1}\right]$$\end{document} is a correction energy coefficient governing how aggressively drift is subtracted. Higher \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{corr}$$\end{document} applies stronger correction and is appropriate at high heterogeneity levels where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.1$$\end{document}, while lower \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{corr}$$\end{document} is used at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=1.0\:$$\end{document}where drift is minimal. The corrected updates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}^{corrected}$$\end{document} from Eq. (47) replace \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}\:$$\end{document}in the cluster aggregation step of Eq. (43), so the complete corrected cluster update becomes as specified in Eq. (48):48\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{g}^{cluster}=\frac{{\sum\:}_{k\in\:{\mathcal{C}}_{g}}{N}_{k}\cdot\:{{\Delta\:}}_{k,t}^{corrected}}{{\sum\:}_{k\in\:{\mathcal{C}}_{g}}{N}_{k}}$$\end{document} Aggregation convergence analysis The convergence behavior of the proposed cluster-weighted aggregation with gradient divergence correction is measured under the standard FL of \:L-smooth and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:$$\end{document}-strongly convex local objectives. The convergence bound for the proposed model after \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{round}$$\end{document} communication rounds are stated as in Eq. (49):49\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbb{E}}\left[ {\parallel \theta _{{T_{{round}} }}^{g} - \theta ^{*} \parallel ^{2} } \right] \le \left( {1 - \mu \eta } \right)^{{T_{{round}} }} \parallel \theta _{0}^{g} - \theta ^{*} \parallel ^{2} + \frac{{\Gamma _{{non - IID}} }}{{T_{{round}} }}$$\end{document} where ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mu\:$$\end{document}’ is the robust convexity constant, ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta$$\end{document}’ is the learning rate from Eq. (19), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{0}^{g}$$\end{document} is the initial global model, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Gamma\:}}_{non-IID}$$\end{document} is a heterogeneity residual term. The heterogeneity residual is defined as in Eq. (50):50\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\Gamma _{{non - IID}} = \sum\limits_{{g = 1}}^{G} {\omega _{g}^{2} \cdot \overline{{{\mathrm{JSD}}}} _{g} \cdot \left( {1 - \mu _{{corr}} } \right)^{2} \cdot G^{2} \tau ^{2} \eta ^{2} }$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\omega\:}_{g}$$\end{document} is the inter-cluster weight from Eq. (45), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overline{{{\mathrm{JSD}}}} _{g}$$\end{document} is the mean cluster divergence, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{corr}$$\end{document} is the correction momentum from Eq. (47). PBFT-secured model aggregation protocol The cluster-weighted aggregation in Sect. 5.4 accepts honest updates, but under the threat model in Sect. 3.4, Byzantine institutions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:B\subset\:K$$\end{document} may submit manipulated updates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widetilde{\Delta }_{{k,t}}$$\end{document} to degrade accuracy or induce targeted misclassification. Without validation, such updates can corrupt the global model and violate the performance objective in Eq. (9). This work extends prior work16 to secure the aggregation process itself, requiring consortium-wide Byzantine-fault-tolerant validation of each update before it is included in cluster aggregation (Eq. (48). Model update as BC transaction Each institution ‘k’ encapsulates its corrected model update \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}^{corrected}$$\end{document} from Eq. (47) into a structured BC transaction before submission to the aggregation server. The transaction is defined as in Eq. (51):51\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{k,t}=\left\{{\hspace{0.17em}}I{D}_{k},{\hspace{0.25em}\hspace{0.05em}}t,{\hspace{0.25em}\hspace{0.05em}}{{\Delta\:}}_{k,t}^{\mathrm{corrected}},{\hspace{0.25em}\hspace{0.05em}}{M}_{k,t},{\hspace{0.25em}\hspace{0.05em}}{\sigma\:}_{k}{\hspace{0.17em}}\right\}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:I{D}_{k}$$\end{document} is the unique institution identifier, ‘t’ is the communication round index, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}^{corrected}$$\end{document} is the gradient update from Eq. (47), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{M}}_{k,t}$$\end{document} is a metadata vector encoding local training statistics, including loss value, and dataset size \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{k}$$\end{document}, and local heterogeneity estimate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{JSD}({P}_{k}\parallel\:P)$$\end{document} from Eq. (16), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{k}$$\end{document} is a cryptographic signature generated using institution \:k‘s private key. The metadata \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{M}}_{k,t}$$\end{document} serves 2 purposes: it provides the aggregation server with sufficient information to compute the cluster affinity matrix in Eq. (41) without accessing raw local data, and it provides the Byzantine filtering method in Sect. 5.5.3 with statistical signals for detecting anomalous update behavior. Each transaction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{T}}_{k,t}$$\end{document} from Eq. (51) is cryptographically hashed and linked to the previous block as specified in Eq. (52):52\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:hash\left({B}_{j}\right)=H\left({B}_{j-1}\parallel\:{\mathcal{T}}_{k,t}\parallel\:nonce\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{B}_{j-1}$$\end{document} is the previous block, \:nonce is the consensus nonce, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H(\cdot\:)$$\end{document} is a secure hash function. PBFT consensus for aggregation validation Submitted transactions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{\mathcal{T}}_{k,t}{\}}_{k=1}^{K}\right.$$\end{document} from Eq. (51) are validated using a three-phase PBFT protocol before the aggregation server is permitted to process any update. The protocol operates across the ‘k’ consortium nodes and tolerates up to \:f{\stackrel{\prime }{\psi\:}}_{t}+\kappa\:\cdot\:{\sigma\:}_{\psi\:,t}\right\}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\prime }{\psi\:}}_{t}\:$$\end{document}and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{\psi\:,t}$$\end{document} are the mean and standard deviation of anomaly scores across all institutions in round ‘t’, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\kappa\:>0$$\end{document} is a sensitivity parameter. Updates from institutions in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{B}}_{suspect}$$\end{document} from Eq. (58) are excluded from the cluster aggregation in Eq. (48), and the corresponding institutions are reported to the SC layer in Sect. 5.5.4. The filtered honest update set admitted to aggregation is defined as in Eq. (59):59\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{K}}_{honest,t}=\mathcal{K}\setminus\:{\mathcal{B}}_{suspect,t}$$\end{document} where all subsequent aggregation operations in Eq. (43) through Eq. (48) are restricted to institutions in \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{K}}_{honest,t}$$\end{document}from Eq. (59). SC-enforced participation policy Each institution’s participation record is continued on the BC ledger as a reputation score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{k}$$\end{document} updated at the end of each communication round as agreed in Eq. (60):60\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho \:_{{k,t + 1}} = \lambda \:_{{\rho \:}} \cdot \:\rho \:_{{k,t}} + (1 - \lambda \:_{{\rho \:}} ) \cdot \:1\left[ {k \notin {\mathcal{B}}_{{suspect,t}} } \right]$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{\rho\:}\in\:\left(\mathrm{0,1}\right)$$\end{document} is a decay factor controlling the memory horizon of the reputation update, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1[\cdot\:]$$\end{document} is the indicator function returning 1 if institution ‘k’ was not flagged as Byzantine-suspect in round ‘t’ and 0 otherwise. An institution whose reputation score falls below a suspension threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{min}$$\end{document} is automatically excluded from future rounds by the SC as agreed in Eq. (61):61\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:permi{t}_{k}=\left\{\begin{array}{cc}1&\:\mathrm{if\:}{\rho\:}_{k,t}\ge\:{\rho\:}_{min}\\\:0&\:\mathrm{otherwise}\end{array}\right.$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:permi{t}_{k}=0$$\end{document} suspends institution ‘k’ from submitting updates until its reputation recovers through sustained honest participation. Global model distribution and local fine-tuning PBFT-validated aggregation had generated an updated global model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document} from Eq. (44). This model captures the view of the honest majority and also corrects non-IID divergence from local updates. Before redistribution, two things need to be handled. First is secure transmission under privacy and integrity constraints, and second is controlled local adaptation that does not break convergence guarantees. These two things are addressed through the secure model broadcast protocol and constrained local fine-tuning procedure. Secure global model broadcast The aggregation server broadcasts the updated global model parameter vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document} to all institutions sustaining \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{permit}}_{k}=1$$\end{document} at the end of each communication round. The broadcast message is encapsulated as a signed BC transaction, as defined in Eq. (62):62\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\begin{array}{cccc}&\:{M}_{\mathrm{broadcast},{\hspace{0.17em}}t+1}=\left\{{\hspace{0.17em}}t+1,{\hspace{0.25em}\hspace{0.05em}}{\theta\:}_{t+1}^{g},{\hspace{0.25em}\hspace{0.05em}}\mathrm{d}\mathrm{i}\mathrm{g}\mathrm{e}\mathrm{s}\mathrm{t}{\hspace{0.17em}}\left({\theta\:}_{t+1}^{g}\right),{\hspace{0.25em}\hspace{0.05em}}{\sigma\:}_{\mathrm{server}}{\hspace{0.17em}}\right\}&\:&\:\end{array}$$\end{document} where \:t+1 is the round index, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document} is the updated global parameter vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:digest\left({\theta\:}_{t+1}^{g}\right)=H\left({\theta\:}_{t+1}^{g}\right)$$\end{document} is the cryptographic hash of the global model, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{server}$$\end{document} is the aggregation server’s digital signature. Upon receipt, each institution ‘k’ verifies the integrity of the received model by recomputing the digest and comparing it against \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:digest\left({\theta\:}_{t+1}^{g}\right)$$\end{document} in Eq. (62). A verification failure at institution ‘k’triggers a re-request through the consortium BC, ensuring that no institution initializes local training from a corrupted or tampered global model. Only upon successful verification does institution ‘k’ update its local model initialization to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document}for the subsequent round of local training per Eq. (19). Constrained local fine-tuning A constrained local fine-tuning method is presented here that limits the magnitude of local parameter updates relative to the received global model, bounding the contribution of each local step to the drift accumulation captured in Eq. (40). The constrained local objective at the institution ‘k’in round \:t+1 augments the standard local objective from Eq. (3) with a proximal regularization term as agreed in Eq. (63):63\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{k}^{prox}\left(\theta\:\right)={F}_{k}\left(\theta\:\right)+\frac{{\mu\:}_{prox}}{2}{\parallel\:\theta\:-{\theta\:}_{t+1}^{g}\parallel\:}^{2}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{k}\left(\theta\:\right)$$\end{document} is the local objective from Eq. (3), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{prox}>0$$\end{document} is the proximal regularization coefficient, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\parallel\:\theta\:-{\theta\:}_{t+1}^{g}\parallel\:}^{2}$$\end{document} penalizes deviation of the locally updated parameters from the global initialization. The proximal term in Eq. (63) acts as a soft constraint that prevents local training from moving the institution’s parameters too far from the global model, directly bounding the drift term in Eq. (47) without requiring communication between institutions during local training. The gradient of Eq. (63) with respect to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document} is specified as in Eq. (64):64\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\nabla \:F_{k}^{{prox}} \left( {\theta \:} \right) = \nabla \:F_{k} \left( {\theta \:} \right) + \mu \:_{{prox}} \left( {\theta - \theta _{{t + 1}}^{g} } \right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\nabla\:{F}_{k}\left(\theta\:\right)$$\end{document} is the standard local gradient and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{prox}(\theta\:-{\theta\:}_{t+1}^{g})\:$$\end{document} is the proximal correction. The local training step from Eq. (19) is updated to use the proximal gradient from Eq. (64) as assumed in Eq. (65):65\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k,t+1}={\theta\:}_{k,t}-\eta\:\nabla\:{F}_{k}^{prox}\left({\theta\:}_{k,t}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k,t+1}$$\end{document} is the locally updated parameter at step \:t+1 within the communication round. Interaction between fine-tuning and non-IID adaptation The constrained fine-tuning interacts with the Non-IID-adapted DSAE in a coordinated manner due to competing objectives. Since DSAE parameters are part of the global model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document}, they are nominally subject to the proximal constraint in Eq. (63), while the alignment regularizer \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Omega\:}}_{\mathrm{align},k}$$\end{document} from Eq. (27) drives local embeddings toward the global distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}^{g}\left(z\right)$$\end{document}, requiring divergence from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document}. This conflict is resolved by selectively applying the proximal constraint only to the detection module parameters—BiLSTM, PoE modules, and the fusion layer—while excluding DSAE parameters, which are optimized solely using the full DSAE objective in Eq. (28), thereby ensuring alignment without compromising convergence. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{det}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{dsae}$$\end{document} are forming completely disjoint parameter sets with no shared weights between them, the proximal regularization operating on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{det}$$\end{document} and the alignment regularizer operating on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{dsae}$$\end{document} are inducing separate parts of the parameter space without introducing optimization conflict or instability. The partitioned local objective is defined as in Eq. (66):66\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{k}^{total}\left(\theta\:\right)={F}_{k}^{prox}\left({\theta\:}_{det}\right)+{L}_{DSAE,k}\left({\theta\:}_{dsae}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{det}$$\end{document} is the detection module parameters subject to the proximal constraint from Eq. (63), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{dsae}$$\end{document} is the DSAE parameters optimized under Eq. (28), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:={\theta\:}_{det}\cup\:{\theta\:}_{dsae}$$\end{document} is the complete local parameter set. Non-IID-aware federated training algorithm The training method of the proposed model unifies local feature extraction (FE), detection engine optimization, non-IID-aware aggregation, Byzantine-resilient validation, and constrained fine-tuning into a single FL loop. The 3 phases operate in sequence within each communication round to coordinate the interaction among clustering, drift correction, and PBFT validation. In the first phase, each institution is completing local training and drift-corrected updates \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Delta\:}}_{k,t}^{corrected}$$\end{document}, from Eq. (47), submitted as PBFT-validated BC transactions. In the second phase, PBFT consensus from Sect. 5.5 validates submitted transactions, and Byzantine filtering from Eq. (57) to Eq. (59) forms the honest institution set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{K}}_{honest,t}$$\end{document}, only updates from this set are acknowledged for aggregation. In the third phase, the pairwise affinity matrix from Eq. (41) is computed using metadata of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{k,t}$$\end{document} from validated transactions, spectral clustering partitions honest institutions into groups \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left\{{C}_{g}\right\}}_{g=1}^{G}$$\end{document}, and cluster-weighted aggregation from Eqs. (44)–(45) generates the updated global model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}^{g}$$\end{document}. So drift correction output feeds into PBFT validation, PBFT output gates cluster aggregation, and cluster aggregation output is redistributed as the next round initialization — forming a tightly coupled three-phase cycle. Algorithm 1 formalizes this technique across all ‘k’ institutions and the aggregation server within each communication round ‘t’ Algorithm 1Non-IID-aware federated intrusion detection training. Non-IID-aware federated intrusion detection training. The algorithm separates institution-parallel local operations from aggregation-server centralized operations, reflecting the physical deployment boundary of the FL. The DSAE pre-training executes only at \:t=0, after which fine-tuning under the full objective from Eq. (28) resumes each round with an updated global reference distribution. The convergence check monitors parameter stability across consecutive global models, terminating training when parameter updates stabilize to within tolerance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\varepsilon}_{conv}$$\end{document}. The BC ledger ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}$$\end{document}’ collects all transaction and round summary records across the full training horizon, providing the complete immutable audit trail required for regulatory compliance in multi-institutional HCS deployments. Datasets, experimental setup and implementation Dataset description Two publicly available datasets are used for federated evaluation, and their statistical characteristics are summarized in Table 1. IoT healthcare security dataset (IoT-Flock) The IoT-Flock dataset14 is directly relevant to MIoT deployment environments and provides a realistic simulation of clinical device communication behavior. The dataset was generated to represent an intensive care unit setting with two patient beds, each with nine dedicated monitoring sensors and a separate bed control unit. Traffic flows from this simulation match the communication behaviors and protocol features of real HCS-IoT deployments and also include periodic, low-latency telemetry patterns specific to medical device traffic rather than those of general IoT environments. Normal and malicious traffic flows are generated using the IoT-Flock traffic generation model. This model supports modern healthcare IoT protocols such as Message Queuing Telemetry Transport (MQTT) and Constrained Application Protocol (CoAP). Benign traffic covers environmental reporting, medication-delivery status updates, and continuous health-metric monitoring streams. Attack traffic encompasses malicious attempts to compromise authentication systems, disrupt communication protocol integrity, and inject data across device channels. The dataset is distributed in Packet Capture (PCAP) format and is converted to CSV-based FS using pcap2csv conversion scripts. And this conversion produces flow-level records with protocol-defined features and healthcare-specific metadata. This dataset contains 1,347 labeled attack events and 4,231 benign tests across 3 attack classes: DDoS, MiTM, and BFA. CICIoT2023 CICIoT202315 was released by the Canadian Institute for Cybersecurity and is derived from a large-scale practical testbed comprising 105 real IoT devices spanning multiple device categories and application domains. The dataset included 33 distinct attack types grouped into 7 types, including reconnaissance and detection evasion, web-based exploitation, BFA, and spoofing, as well as 2 variants of Mirai traffic generation. This dataset included tools such as Nmap for network reconnaissance, Hping3 for volumetric flooding, and Metasploit for multi-stage exploitation. Each network flow record has 46 FEs that cover packet-level descriptors, flow-level statistics, and behavioral indicators. Ground-truth labels in this dataset support binary anomaly detection (AD), and multi-class attack classification, together with attack-type-specific evaluation tasks. This dataset has more than 47 million labeled flow records. Both raw network captures and pre-processed FS versions of this dataset are publicly available. This study uses the FS version, which maintains an input space of 46 dimensions across all test cases. Since the proposed model is built around 3 healthcare IoT attack classes, a label mapping procedure is applied to the 33 original CICIoT2023 attack types before partitioning. Volumetric flooding variants, such as UDP flood, ICMP flood, SYN flood, and HTTP flood, as well as Slowloris, are mapped to the DDoS class. Spoofing variants, such as ARP and DNS spoofing, are mapped to the MiTM class because these attacks operate through interception mechanisms that are consistent with MiTM behavior in healthcare IoT protocols. BFA is mapped directly to the BFA. Reconnaissance and web-based exploitation, along with Mirai-based botnet types, do not represent attack types present in the simulated HCS-IoT deployment, and these flows are excluded. This mapping generates the test partition reported in Table 1, comprising 4,892 attack records and 8,756 benign records across the 3 target classes. Table 1Dataset features.PropertyIoT-FlockCICIoT2023Attack Records13474,892 (evaluation partition)Benign Records42318,756 (evaluation partition)Total Flow Records188,69447,000,000+Attack Classes3 (DDoS, MiTM, BFA)33 across 7 categoriesFeaturesProtocol-defined flow features46 extracted flow featuresPrimary ProtocolsMQTT, CoAPMixed multi-vendorSource EnvironmentSimulated ICU (2 beds, 18 sensors)105 real IoT devicesFormatPCAP converted to CSVRaw PCAP and feature-selected CSVAvailabilityPublic (IEEE DataPort)Public (Canadian Institute for Cybersecurity) Table 1Dataset features.PropertyIoT-FlockCICIoT2023Attack Records13474,892 (evaluation partition)Benign Records42318,756 (evaluation partition)Total Flow Records188,69447,000,000+Attack Classes3 (DDoS, MiTM, BFA)33 across 7 categoriesFeaturesProtocol-defined flow features46 extracted flow featuresPrimary ProtocolsMQTT, CoAPMixed multi-vendorSource EnvironmentSimulated ICU (2 beds, 18 sensors)105 real IoT devicesFormatPCAP converted to CSVRaw PCAP and feature-selected CSVAvailabilityPublic (IEEE DataPort)Public (Canadian Institute for Cybersecurity) Dataset features. Federated node configuration and institutional simulation The federated evaluation requires a simulated multi-institutional network whose node configuration reflects the heterogeneous administrative network of real healthcare IoT consortia. Since real distributed patient data across multiple institutions is unavailable, the federation is simulated by partitioning the IoT-Flock and CICIoT2023 across \:K=12 virtual institutional nodes using the Flower (flwr) FL model. The 12 nodes in this study were grouped into 4 practical groups based on their clinical roles within the HCS. Primary care nodes encompass 4 institutions, all large general hospitals with broad device inventories and high traffic volumes. These nodes experience diverse attack distributions covering DDoS, MiTM, and BFA. The 3 institutions are grouped under specialty care nodes, and these services focus on clinical areas such as cardiology or oncology. These nodes have narrower device types and higher wireless device density, leading to greater MiTM exposure than other node types. The network operations nodes comprise two institutions that provide centralized monitoring and infrastructure management. Attack-class exposure in these nodes is nearly uniform across all three categories. Administrative and research nodes span 3 institutions, including outpatient clinics, health research centers, and regional administrative offices. Traffic through these nodes is mostly benign, and exposure to attacks is limited. IoT-Flock samples are distributed across the 4 primary care nodes and 2 specialty care nodes, exploiting the dataset’s intensive care unit (ICU) bed-sensor-control unit groupings as the natural partitioning basis. CICIoT2023 samples are distributed across all 12 nodes by device type, with each node receiving traffic from a subset of the 105 device types, weighted by their functional group profiles. The protocol mixture vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\pi\:}_{k}$$\end{document} from Eq. (17), the value is assigned per node, as summarized in Table 2. Table 2Federated node configuration.Node GroupCountDataset SourceDominant ProtocolsAttack PriorPrimary Care4IoT-Flock + CICIoT2023MQTT, CoAPDDoS, BFASpecialty Care3IoT-Flock + CICIoT2023MQTT, HL7-FHIRMiTM elevatedNetwork Operations2CICIoT2023Mixed multi-protocolUniform across classesAdministrative/Research3CICIoT2023CoAP, DICOMBenign dominant Table 2Federated node configuration.Node GroupCountDataset SourceDominant ProtocolsAttack PriorPrimary Care4IoT-Flock + CICIoT2023MQTT, CoAPDDoS, BFASpecialty Care3IoT-Flock + CICIoT2023MQTT, HL7-FHIRMiTM elevatedNetwork Operations2CICIoT2023Mixed multi-protocolUniform across classesAdministrative/Research3CICIoT2023CoAP, DICOMBenign dominant Federated node configuration. Non-IID partition implementation The non-IID partitioning method instantiates the formal models derived in Sect. 4 into a concrete data distribution across the 12 simulated nodes. All partitioning operations use a fixed random seed of 42, consistently applied across Python’s random module, NumPy, and the Flower network, ensuring full reproducibility across experimental runs. The 3 partitioning dimensions are implemented autonomously and in combination, corresponding to the quantity skew, label skew, and feature skew models described in Sect. 4.2–4.4. Quantity skew is implemented by sampling node dataset sizes from the Dirichlet distribution in Eq. (13) with concentration parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:\in\:\:\in\:\:\{0.3,\:0.7,\:1.0\},\:thereby$$\end{document} generating high, moderate, and low quantity imbalance conditions. At \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:=0.3$$\end{document}, the largest node holds approximately 40% of total samples, while the smallest holds under 2%, generating a Gini coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{N}$$\end{document} from Eq. (14) of nearly 0.61. At \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:=1.0$$\end{document}, node sizes are near-uniform with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{N}\approx\:0.12$$\end{document}. Label skew is implemented using the Dirichlet-based class proportion sampling from Eq. (7) at 3 concentration levels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:\in\:\left\{\mathrm{0.1,0.5,1.0}\right\}$$\end{document}. Group-level prior vectors \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{p}}^{\left(g\right)}$$\end{document}from Eq. (15) are constructed per the functional group assignments in Table 2, with temperature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{prior}=0.5$$\end{document} generating moderately peaked group priors. The mean Jensen-Shannon divergence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\overline{{{\mathrm{JSD}}}}$$\end{document} from Eq. (16) is computed after partitioning and reported per ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}$$\end{document}’ level in Sect. 7.3 to quantify achieved heterogeneity. Feature skew is implemented by assigning protocol mixture weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\pi\:}_{k}$$\end{document} from Eq. (17) per the node group assignments in Table 2, with MQTT-dominant nodes receiving \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\pi\:}_{k,MQTT}=0.7$$\end{document} and mixed-protocol nodes receiving uniform weights across all four protocol types. The combined heterogeneity condition applies to all 3 skew types simultaneously at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:=0.7$$\end{document}, and protocol mixture weights per Table 2. Baseline methods for comparison Six baseline methods are selected for performance comparison, spanning centralized DL, standard federated averaging, and heterogeneity-aware federated variants. The centralized hybrid BC + DL16 serves as the upper-bound reference, trained on pooled data from all institutions without federated constraints. Standard FedAvg serves as the foundational federated baseline, isolating the contribution of the non-IID-aware cluster-weighted mechanism over naive data-proportional averaging. FedProx extends FedAvg with the proximal regularization but without cluster-weighted aggregation. SCAFFOLD, a variance-reduction FL applying control variates to correct client drift, represents the current state-of-the-art (SOTA) in non-IID-FL. FedNova, a normalized averaging method that accounts for heterogeneous local update magnitudes, is included as a comparison with the cluster-weighting method. A federated variant network without PBFT-secured aggregation is included to isolate the definite contribution of the Byzantine-resilient gradient filtering. Hardware and software implementation details Hardware and software configuration Training and aggregation experiments are run on an NVIDIA DGX-A100 server with 8 × A100 GPUs with 80 GB of memory. Edge-level detection latency is measured on an NVIDIA Jetson AGX Orin, simulating the perception-layer deployment environment. Each of the 12 federated nodes is simulated as a separate process on the DGX-A100, with inter-process communication emulating the network transmission overhead of a real distributed deployment. The complete software stack comprises PyTorch 2.1.0 with CUDA 12.1 for DL components, Flower (flwr 1.6.0) for FL setup, and Hyperledger Fabric 2.5.4 with PBFT consensus for BC simulation. SC logic is implemented in the Go chaincode. Network configuration The network configurations for all model components are summarized in Table 3. The DSAE encoder input dimension of 87 corresponds to the feature space of IoT-Flock, which had 87 protocol-defined flow features after pcap2csv conversion. For CICIoT2023, which had 46 extracted flow features, zero-padding is applied to extend each input vector to 87 dimensions, maintaining a unified network across both datasets without requiring separate model instances per dataset. Table 3Model network configuration.ComponentParameterValueDSAEEncoder dimensions[87, 64, 32, 16, 8]DSAEDecoder dimensions[8, 16, 32, 64, 87]DSAESparsity target ρ‘’0.05DSAESparsity coefficient β‘’3.0DSAEBottleneck dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{L}$$\end{document}‘’8BiLSTMHidden dimension ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{h}$$\end{document}’ per direction64BiLSTMCombined hidden state dimension128BiLSTMSequence length ‘\:T’10RF (DDoS)Number of trees ‘\:M’100RF (DDoS)Maximum depth10MLP (DDoS)Hidden layer dimensions[128, 64, 1]DQN (MiTM)Hidden layer dimensions[128, 128]DQN (MiTM)Replay buffer size10,000CNN (Brute-Force)Conv2D filters (layer 1)32CNN (Brute-Force)Conv2D filters (layer 2)64CNN (Brute-Force)Pooling window2 × 2Spectral ClusteringNumber of clusters ‘\:G’4 Table 3Model network configuration.ComponentParameterValueDSAEEncoder dimensions[87, 64, 32, 16, 8]DSAEDecoder dimensions[8, 16, 32, 64, 87]DSAESparsity target ρ‘’0.05DSAESparsity coefficient β‘’3.0DSAEBottleneck dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{L}$$\end{document}‘’8BiLSTMHidden dimension ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{h}$$\end{document}’ per direction64BiLSTMCombined hidden state dimension128BiLSTMSequence length ‘\:T’10RF (DDoS)Number of trees ‘\:M’100RF (DDoS)Maximum depth10MLP (DDoS)Hidden layer dimensions[128, 64, 1]DQN (MiTM)Hidden layer dimensions[128, 128]DQN (MiTM)Replay buffer size10,000CNN (Brute-Force)Conv2D filters (layer 1)32CNN (Brute-Force)Conv2D filters (layer 2)64CNN (Brute-Force)Pooling window2 × 2Spectral ClusteringNumber of clusters ‘\:G’4 Model network configuration. The DSAE input dimension of 87 corresponds to the IoT-Flock feature space. CICIoT2023 inputs (46-dim) are zero-padded to 87 dimensions to maintain a unified network across both datasets. Federated training hyperparameters All hyperparameters governing the federated training method of Algorithm 1 are consolidated in Table 4. Table 4Federated training hyperparameters.ParameterSymbolValueLocal learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document} 0.001Local training steps per round \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document} 10Adam decay rate 1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{1}$$\end{document} 0.9Adam decay rate 2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{2}$$\end{document} 0.999Weight decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:{10}^{-4}$$\end{document} Proximal coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{prox}$$\end{document} 0.01Alignment coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{align}$$\end{document} 0.1Correction momentum \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{corr}$$\end{document} 0.5Sparsity regularization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} 3.0Byzantine filtering sensitivity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\kappa\:$$\end{document} 2.0Reputation decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{\rho\:}$$\end{document} 0.9Suspension threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{min}$$\end{document} 0.4Affinity bandwidth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{aff}^{2}$$\end{document} 1.0Cluster temperature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{clust}$$\end{document} 1.0DSAE pre-training epochs–100BiLSTM training epochs–50Expert module training epochs–30Maximum federation rounds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{max}$$\end{document} 200Convergence tolerance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varepsilon$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:{10}^{-4}$$\end{document} Label skew levels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} {0.1, 0.5, 1.0}Quantity skew levels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{qty}$$\end{document} {0.3, 0.7, 1.0} Table 4Federated training hyperparameters.ParameterSymbolValueLocal learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document} 0.001Local training steps per round \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document} 10Adam decay rate 1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{1}$$\end{document} 0.9Adam decay rate 2 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{2}$$\end{document} 0.999Weight decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:{10}^{-4}$$\end{document} Proximal coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{prox}$$\end{document} 0.01Alignment coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{align}$$\end{document} 0.1Correction momentum \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{corr}$$\end{document} 0.5Sparsity regularization \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} 3.0Byzantine filtering sensitivity \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\kappa\:$$\end{document} 2.0Reputation decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\lambda\:}_{\rho\:}$$\end{document} 0.9Suspension threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{min}$$\end{document} 0.4Affinity bandwidth \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{aff}^{2}$$\end{document} 1.0Cluster temperature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{clust}$$\end{document} 1.0DSAE pre-training epochs–100BiLSTM training epochs–50Expert module training epochs–30Maximum federation rounds \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{max}$$\end{document} 200Convergence tolerance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\varepsilon$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:1\times\:{10}^{-4}$$\end{document} Label skew levels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} {0.1, 0.5, 1.0}Quantity skew levels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\beta\:}_{qty}$$\end{document} {0.3, 0.7, 1.0} Federated training hyperparameters. Computational complexity analysis The per-round computational complexity and communication cost of the proposed model are compared with those of all baseline methods in Table 5. Complexity is stated in terms of model parameter count ‘p’, number of institutions ‘k’, and local dataset size ‘\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{N}_{k}$$\end{document}’, local training steps ‘τ’, and number of clusters ‘g’. All federated methods share the same base local computation cost of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} per institution per round. The proposed model adds a DSAE training cost of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left({N}_{k}{d}_{0}{d}_{L}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{0}=87$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{L}=8$$\end{document}, which is small relative to the gradient cost of the detection module. At the aggregation server, FedAvg, FedProx, and FedNova are operating at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)\:$$\end{document}while the proposed model adds spectral clustering at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}({K}^{2}+{K}^{1.5})$$\end{document}, Byzantine filtering at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document}, and PBFT consensus at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left({K}^{2}\right)$$\end{document}. Since \:K=12 is fixed and small, these overheads are not significant relative to gradient computation. For communication, all methods transmit \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document} per round, and the proposed model adds only a fixed-size metadata vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}(C+2)$$\end{document} per institution. Table 5Per-round computational complexity and communication cost.MethodLocal computationAggregation costCommunicationAdditional OverheadFedAvg \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document} NoneFedProx \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document} Proximal term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document}SCAFFOLD \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(2p\right)$$\end{document} Control variate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document}FedNova \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document} Normalized update \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document}Proposed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}(\tau\:{N}_{k}p+{N}_{k}{d}_{0}{d}_{L})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}(Kp+{K}^{2})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}(p+C)$$\end{document} PBFT \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left({K}^{2}\right)$$\end{document}, clustering \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left({K}^{1.5}\right)$$\end{document}, filtering \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} Table 5Per-round computational complexity and communication cost.MethodLocal computationAggregation costCommunicationAdditional OverheadFedAvg \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document} NoneFedProx \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document} Proximal term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document}SCAFFOLD \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(2p\right)$$\end{document} Control variate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document}FedNova \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(\tau\:{N}_{k}p\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document} Normalized update \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(p\right)$$\end{document}Proposed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}(\tau\:{N}_{k}p+{N}_{k}{d}_{0}{d}_{L})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}(Kp+{K}^{2})$$\end{document} \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}(p+C)$$\end{document} PBFT \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left({K}^{2}\right)$$\end{document}, clustering \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left({K}^{1.5}\right)$$\end{document}, filtering \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}\left(Kp\right)$$\end{document} Per-round computational complexity and communication cost. Results and analysis Detection performance under IID conditions Table 6; Fig. 3 present the detection performance of all 6 baselines alongside the proposed model under an IID data distribution. Table 6Detection performance under IID conditions.MethodIoT-Flock Accuracy (%)IoT-Flock F1 (%)IoT-Flock AUC-ROCCICIoT2023 Accuracy (%)CICIoT2023 F1 (%)CICIoT2023 AUC-ROCCentralized BC + DL196.7392.740.983493.5885.310.9456FedAvg95.8491.230.976192.4783.940.9378FedProx96.1291.870.978992.8384.520.9401SCAFFOLD95.6790.940.974392.1483.270.9352FedNova95.9391.460.977292.6184.130.9389Federated BC + DL w/o PBFT95.4790.730.972491.8983.060.9331Proposed Model96.4192.180.981293.2484.970.9438 Table 6Detection performance under IID conditions.MethodIoT-Flock Accuracy (%)IoT-Flock F1 (%)IoT-Flock AUC-ROCCICIoT2023 Accuracy (%)CICIoT2023 F1 (%)CICIoT2023 AUC-ROCCentralized BC + DL196.7392.740.983493.5885.310.9456FedAvg95.8491.230.976192.4783.940.9378FedProx96.1291.870.978992.8384.520.9401SCAFFOLD95.6790.940.974392.1483.270.9352FedNova95.9391.460.977292.6184.130.9389Federated BC + DL w/o PBFT95.4790.730.972491.8983.060.9331Proposed Model96.4192.180.981293.2484.970.9438 Detection performance under IID conditions. Under IID conditions, all federated methods perform close to the centralized upper bound, with accuracy gaps ranging from 0.32% to 1.26% on IoT-Flock and from 0.11% to 1.69% on CICIoT2023. The proposed model achieved 96.41% on IoT-Flock and 93.24% on CICIoT2023, which are 0.32% and 0.34% below the centralized model. FedAvg and FedProx also perform well under IID conditions and trail the proposed model by 0.57% and 0.29% on IoT-Flock. SCAFFOLD and FedNova show similar IID performances. Federated variant without PBFT performs at a similar level to other federated baselines under IID conditions. SVM achieves 87.65%, and Random Forest (RF) achieves 89.34% on IoT-Flock under IID conditions. The proposed model outperforms RF by 7.07% on IoT-Flock and 6.52% on CICIoT2023. Fig. 3Detection accuracy (%) under IID conditions. Detection accuracy (%) under IID conditions. Detection performance under non-IID conditions Table 7; Fig. 4 present detection accuracy under quantity skew, label skew, feature skew, and combined skew conditions across all federated methods on both datasets. Table 7Detection accuracy (%) under non-IID conditions.MethodQuantity Skew β = 0.3Quantity Skew β = 0.7Label Skew α = 0.1Label Skew α = 0.5Feature SkewCombinedIoT-Flock FedAvg91.23 ± 1.8493.47 ± 1.2374.83 ± 3.4788.64 ± 1.5687.92 ± 1.6772.38 ± 3.91 FedProx92.84 ± 1.4394.13 ± 0.9779.47 ± 2.8390.23 ± 1.2489.34 ± 1.3876.94 ± 3.14 SCAFFOLD90.67 ± 2.1392.83 ± 1.4731.24 ± 8.6384.17 ± 2.3486.73 ± 1.9428.83 ± 9.17 FedNova92.13 ± 1.6193.84 ± 1.0478.93 ± 2.9489.74 ± 1.3388.67 ± 1.5275.47 ± 3.28 Fed. BC + DL w/o PBFT91.84 ± 1.7393.26 ± 1.1877.42 ± 3.1289.13 ± 1.4187.84 ± 1.5974.83 ± 3.47 Proposed Model94.73 ± 0.8395.84 ± 0.6191.47 ± 1.2494.82 ± 0.7493.26 ± 0.8990.83 ± 1.37CICIoT2023 FedAvg87.34 ± 2.1389.83 ± 1.4771.24 ± 3.8384.73 ± 1.8483.47 ± 1.9368.93 ± 4.23 FedProx88.93 ± 1.6790.47 ± 1.1375.83 ± 3.1486.24 ± 1.4384.93 ± 1.6273.47 ± 3.57 SCAFFOLD86.73 ± 2.4789.14 ± 1.6319.47 ± 9.3480.83 ± 2.6782.34 ± 2.1317.83 ± 9.84 FedNova88.47 ± 1.8390.13 ± 1.2474.93 ± 3.2785.83 ± 1.5784.17 ± 1.7472.83 ± 3.73 Fed. BC + DL w/o PBFT87.93 ± 1.9489.67 ± 1.3673.84 ± 3.4385.34 ± 1.6383.73 ± 1.8771.94 ± 3.84 Proposed Model91.24 ± 0.9492.47 ± 0.7387.83 ± 1.4791.34 ± 0.8789.73 ± 1.0386.47 ± 1.53± values represent standard deviation across α ∈ {0.1, 0.5, 1.0} heterogeneity severity levels. Table 7Detection accuracy (%) under non-IID conditions.MethodQuantity Skew β = 0.3Quantity Skew β = 0.7Label Skew α = 0.1Label Skew α = 0.5Feature SkewCombinedIoT-Flock FedAvg91.23 ± 1.8493.47 ± 1.2374.83 ± 3.4788.64 ± 1.5687.92 ± 1.6772.38 ± 3.91 FedProx92.84 ± 1.4394.13 ± 0.9779.47 ± 2.8390.23 ± 1.2489.34 ± 1.3876.94 ± 3.14 SCAFFOLD90.67 ± 2.1392.83 ± 1.4731.24 ± 8.6384.17 ± 2.3486.73 ± 1.9428.83 ± 9.17 FedNova92.13 ± 1.6193.84 ± 1.0478.93 ± 2.9489.74 ± 1.3388.67 ± 1.5275.47 ± 3.28 Fed. BC + DL w/o PBFT91.84 ± 1.7393.26 ± 1.1877.42 ± 3.1289.13 ± 1.4187.84 ± 1.5974.83 ± 3.47 Proposed Model94.73 ± 0.8395.84 ± 0.6191.47 ± 1.2494.82 ± 0.7493.26 ± 0.8990.83 ± 1.37CICIoT2023 FedAvg87.34 ± 2.1389.83 ± 1.4771.24 ± 3.8384.73 ± 1.8483.47 ± 1.9368.93 ± 4.23 FedProx88.93 ± 1.6790.47 ± 1.1375.83 ± 3.1486.24 ± 1.4384.93 ± 1.6273.47 ± 3.57 SCAFFOLD86.73 ± 2.4789.14 ± 1.6319.47 ± 9.3480.83 ± 2.6782.34 ± 2.1317.83 ± 9.84 FedNova88.47 ± 1.8390.13 ± 1.2474.93 ± 3.2785.83 ± 1.5784.17 ± 1.7472.83 ± 3.73 Fed. BC + DL w/o PBFT87.93 ± 1.9489.67 ± 1.3673.84 ± 3.4385.34 ± 1.6383.73 ± 1.8771.94 ± 3.84 Proposed Model91.24 ± 0.9492.47 ± 0.7387.83 ± 1.4791.34 ± 0.8789.73 ± 1.0386.47 ± 1.53± values represent standard deviation across α ∈ {0.1, 0.5, 1.0} heterogeneity severity levels. Detection accuracy (%) under non-IID conditions. ± values represent standard deviation across α ∈ {0.1, 0.5, 1.0} heterogeneity severity levels. Under quantity skew at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:=0.3,\:$$\end{document} the proposed model achieved 94.73% and 91.24% on IoT-Flock and CICIoT2023, which are 1.99% and 2.34% below the centralized ceiling and 3.50% and 3.90% above FedAvg. At \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:=0.7,$$\end{document} the gap over FedAvg is 2.37% and 2.64%. FedProx and FedNova are within 0.71% and 0.60% of each other across both quantity skew levels on IoT-Flock. The proposed model is 1.71% to 2.00% ahead of both on IoT-Flock. Under label skew at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.1,$$\end{document} performance separation across methods is larger than under other conditions. FedAvg is at 74.83% and 71.24% on IoT-Flock and CICIoT2023, respectively, under these conditions. SCAFFOLD is at 31.24% and 19.47% on IoT-Flock and CICIoT2023. FedProx and FedNova are getting 79.47% and 78.93% on IoT-Flock at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.1$$\end{document}. The proposed model is at 91.47% and 87.83% under this condition. And the performance is 12.00% and 12.00% above FedProx on IoT-Flock and CICIoT2023. At a moderate level \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.5,\:$$\end{document} the proposed model is at 94.82% and 91.34%, which is 1.91% and 2.24% below the centralized ceiling. And FedAvg is 8.09% and 8.85% below the centralized ceiling at this condition. Under feature skew, the proposed model achieves 93.26% and 89.73% on IoT-Flock and CICIoT2023, which are 5.34% and 6.26% above FedAvg. FedProx and FedNova are at 89.34% and 84.93% on CICIoT2023 under feature skew, and their performance is 4.80% and 5.56% below the proposed model. Under the combined condition, the proposed model achieves 90.83% on IoT-Flock and 86.47% on CICIoT2023, which are 5.90% and 7.11% below the centralized ceiling. FedAvg is at 72.38% and 68.93%, which are 18.45% and 17.54% below the proposed model. FedProx and FedNova are at 76.94% and 73.47% on CICIoT2023 under the combined condition, and their performance is 13.89% and 13.00% lower than that of the proposed model. SCAFFOLD is at 28.83% and 17.83% under the combined skew. The Federated BC + DL variant without PBFT is within 0.42% to 2.59% of FedProx across all non-IID conditions on both datasets. And under combined conditions, this variant is 14.00% to 18.89% below the proposed model. Table 8 presents per-class F1-scores under the combined non-IID condition at α = 0.5 and β = 0.7 for the proposed model and the 2 robust baselines. Table 8Per-class F1-score (%) under combined non-IID condition.MethodIoT-FlockCICIoT2023DDoSMiTMBFADDoSMiTMBFAFedAvg74.8368.4771.2371.3464.8369.47FedProx79.3473.2676.8375.8369.4773.84Proposed Model92.4788.3490.8388.7383.4786.94 Table 8Per-class F1-score (%) under combined non-IID condition.MethodIoT-FlockCICIoT2023DDoSMiTMBFADDoSMiTMBFAFedAvg74.8368.4771.2371.3464.8369.47FedProx79.3473.2676.8375.8369.4773.84Proposed Model92.4788.3490.8388.7383.4786.94 Per-class F1-score (%) under combined non-IID condition. Per-class F1 results in Table 8 show that the MiTM class is consistently generating the lowest F1 across all methods on both datasets, which is attributed to its sequential adversarial nature and lower sample representation under the combined non-IID partition. The proposed model achieves 88.34% and 83.47% MiTM F1 scores on IoT-Flock and CICIoT2023, which are 15.08% and 14.00% higher than FedProx on the same classes. The inter-class F1 gap of the proposed model is 4.13% on IoT-Flock and 5.26% on CICIoT2023, and this is smaller than the FedAvg gap of 6.36% and 6.51% and the FedProx gap of 6.08% and 6.36%, signifying that the protocol-aware DSAE normalization and cluster-weighted aggregation are contributing to more balanced detection across all 3 attack types under heterogeneous conditions. Fig. 4Detection accuracy (%) under non-IID conditions. Detection accuracy (%) under non-IID conditions. Non-IID severity and aggregation mechanism analysis Table 9 consolidates the distributional heterogeneity achieved after partitioning at each label skew level, along with detection accuracy and the number of convergence rounds per aggregation method. At \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.1$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\: \:\overline{{{\mathrm{JSD}}}} =0.71$$\end{document}, FedAvg and SCAFFOLD had not converged within 200 rounds on both datasets. FedProx, FedNova, and the federated BC + DL variant without PBFT are at 187–201 rounds on IoT-Flock and CICIoT2023, with accuracies below 80% on IoT-Flock and below 76% on CICIoT2023. The proposed model achieves 147 rounds on IoT-Flock and 163 rounds on CICIoT2023 under these conditions, with 91.47% and 87.83% accuracy. The point at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.5\:$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\: \:\overline{{{\mathrm{JSD}}}} =0.38$$\end{document}, all methods converge within the round budget, as indicated here. The proposed model is at 89 rounds on IoT-Flock and 97 rounds on CICIoT2023, which is 54 and 59 fewer rounds than FedProx and 74 and 81 fewer rounds than FedAvg. FedNova and FedProx converge 37 and 29 rounds later than the proposed model on IoT-Flock at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.5$$\end{document}, respectively. And their final accuracies are 89.74% and 90.23%, compared with 94.82% for the proposed model. The differences in convergence rounds across methods are smaller in this context. Under these conditions, the proposed model is at 41 rounds on IoT-Flock and 48 rounds on CICIoT2023. And accuracy at this condition is 96.41% and 93.24%, which are 0.32% and 0.34% below the centralized ceiling. FedProx is at 54 and 61 rounds on IoT-Flock and CICIoT2023, and FedAvg is at 67 and 74 rounds. Figure 5 shows convergence curves of global model accuracy across federation rounds at three levels for all methods on IoT-Flock. Table 9Non-IID severity, detection accuracy, and convergence rounds.Methodα = 0.1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{{\prime }}{\mathrm{JSD}}$$\end{document} = 0.71α = 0.5 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{{\prime }}{\mathrm{JSD}}$$\end{document} = 0.38α = 1.0 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{{\prime }}{\mathrm{JSD}}$$\end{document} = 0.09Acc (%)RoundsAcc (%)RoundsAcc (%)RoundsIoT-Flock FedAvg74.83 ± 3.47–88.64 ± 1.5614394.47 ± 0.8367 FedProx79.47 ± 2.8318790.23 ± 1.2411895.73 ± 0.6754 SCAFFOLD31.24 ± 8.63–84.17 ± 2.3416495.14 ± 0.7461 FedNova78.93 ± 2.9419389.74 ± 1.3312695.38 ± 0.7158 Fed. BC + DL w/o PBFT77.42 ± 3.1219689.13 ± 1.4113194.83 ± 0.7963 Proposed Model91.47 ± 1.2414794.82 ± 0.748996.41 ± 0.4341CICIoT2023 FedAvg71.24 ± 3.83–84.73 ± 1.8415691.83 ± 0.9474 FedProx75.83 ± 3.1419486.24 ± 1.4312992.47 ± 0.7861 SCAFFOLD19.47 ± 9.34–80.83 ± 2.6717891.23 ± 0.8768 FedNova74.93 ± 3.2719885.83 ± 1.5713792.13 ± 0.8364 Fed. BC + DL w/o PBFT73.84 ± 3.4320185.34 ± 1.6314391.74 ± 0.9171 Proposed Model87.83 ± 1.4716391.34 ± 0.879793.24 ± 0.5148± values represent standard deviation across three heterogeneity severity levels α ∈ {0.1, 0.5, 1.0}. Table 9Non-IID severity, detection accuracy, and convergence rounds.Methodα = 0.1 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{{\prime }}{\mathrm{JSD}}$$\end{document} = 0.71α = 0.5 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{{\prime }}{\mathrm{JSD}}$$\end{document} = 0.38α = 1.0 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{{\prime }}{\mathrm{JSD}}$$\end{document} = 0.09Acc (%)RoundsAcc (%)RoundsAcc (%)RoundsIoT-Flock FedAvg74.83 ± 3.47–88.64 ± 1.5614394.47 ± 0.8367 FedProx79.47 ± 2.8318790.23 ± 1.2411895.73 ± 0.6754 SCAFFOLD31.24 ± 8.63–84.17 ± 2.3416495.14 ± 0.7461 FedNova78.93 ± 2.9419389.74 ± 1.3312695.38 ± 0.7158 Fed. BC + DL w/o PBFT77.42 ± 3.1219689.13 ± 1.4113194.83 ± 0.7963 Proposed Model91.47 ± 1.2414794.82 ± 0.748996.41 ± 0.4341CICIoT2023 FedAvg71.24 ± 3.83–84.73 ± 1.8415691.83 ± 0.9474 FedProx75.83 ± 3.1419486.24 ± 1.4312992.47 ± 0.7861 SCAFFOLD19.47 ± 9.34–80.83 ± 2.6717891.23 ± 0.8768 FedNova74.93 ± 3.2719885.83 ± 1.5713792.13 ± 0.8364 Fed. BC + DL w/o PBFT73.84 ± 3.4320185.34 ± 1.6314391.74 ± 0.9171 Proposed Model87.83 ± 1.4716391.34 ± 0.879793.24 ± 0.5148± values represent standard deviation across three heterogeneity severity levels α ∈ {0.1, 0.5, 1.0}. Non-IID severity, detection accuracy, and convergence rounds. ± values represent standard deviation across three heterogeneity severity levels α ∈ {0.1, 0.5, 1.0}. Fig. 5Convergence plot on IoT-Flock dataset. Convergence plot on IoT-Flock dataset. PBFT consensus and byzantine resilience Table 10 reports BC operational metrics under the 12-node Hyperledger Fabric 2.5.4 PBFT consortium across 5 load scenarios, and Table 9 reports global model accuracy degradation under increasing Byzantine node counts. Table 10BC operational metrics.Load ScenarioThroughput (TPS)Consensus Latency (CL) (ms)SC Execution Time (ms)Validation Success Rate (VSR) (%)Low Load (IoT-Flock)156234.723.4–78.299.91Moderate Load318287.431.7–104.699.84High Load (CICIoT2023)542356.245.8–156.999.73Peak Load (Both Datasets)698423.867.3–203.499.58Emergency Broadcast289198.334.1–46.899.96Multi-Attack Scenario734467.278.2–245.399.41 Table 10BC operational metrics.Load ScenarioThroughput (TPS)Consensus Latency (CL) (ms)SC Execution Time (ms)Validation Success Rate (VSR) (%)Low Load (IoT-Flock)156234.723.4–78.299.91Moderate Load318287.431.7–104.699.84High Load (CICIoT2023)542356.245.8–156.999.73Peak Load (Both Datasets)698423.867.3–203.499.58Emergency Broadcast289198.334.1–46.899.96Multi-Attack Scenario734467.278.2–245.399.41 BC operational metrics. Table 11Global model accuracy under byzantine node counts.Byzantine Nodes \:fHonest MajorityIoT-Flock Accuracy (%)IoT-Flock Accuracy Drop (%)CICIoT2023 Accuracy (%)CICIoT2023 Accuracy Drop (%)\:f=0 (baseline)12/1290.83–86.47– \:f=1 11/1290.470.3686.130.34 \:f=2 10/1289.831.0085.471.00 \:f=3 9/1288.941.8984.731.74 \:f=4 8/1287.833.0083.612.86\:f=4w/o PBFT8/1271.3419.4967.8318.64 Table 11Global model accuracy under byzantine node counts.Byzantine Nodes \:fHonest MajorityIoT-Flock Accuracy (%)IoT-Flock Accuracy Drop (%)CICIoT2023 Accuracy (%)CICIoT2023 Accuracy Drop (%)\:f=0 (baseline)12/1290.83–86.47– \:f=1 11/1290.470.3686.130.34 \:f=2 10/1289.831.0085.471.00 \:f=3 9/1288.941.8984.731.74 \:f=4 8/1287.833.0083.612.86\:f=4w/o PBFT8/1271.3419.4967.8318.64 Global model accuracy under byzantine node counts. Network throughput across 5 load scenarios in Table 10 ranged from 156 TPS at low load to 734 TPS at the multi-attack scenario peak, and CL ranged from 234.7 to 467.2 ms across the same range. All load conditions are maintaining VSR above 99.41%. SC execution times are from 23.4 to 245.3 ms, depending on operation type. Role verification completes in 23.4 ms on IoT-Flock. Audit trail retrieval is at 245.3 ms on CICIoT2023. The emergency broadcast scenario has a VSR of 99.96% and a CL of 198.3 ms. At peak combined load of 698 TPS, CL is 423.8 ms. Table 11 shows the Byzantine resilience contribution of PBFT-secured aggregation by comparing accuracy with and without a filtering mechanism under \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\in\:\left\{\mathrm{1,2},\mathrm{3,4}\right\}\:$$\end{document}malicious nodes. Without PBFT filtering, 4 Byzantine nodes reduced the combined-condition accuracy from 90.83% to 71.34% on IoT-Flock and from 86.47% to 67.83% on CICIoT2023, resulting in drops of 19.49% and 18.64%. With PBFT-secured filtering active, the same 4 Byzantine nodes are bringing accuracy down by only 3.00% and 2.86% on IoT-Flock and CICIoT2023, and the global model is at 87.83% and 83.61%. At \:f=1 and \:f=2, the accuracy drops on IoT-Flock are 0.36% and 1.00%. The reputation decay mechanism from Sect. 5.5.4 prevents persistent Byzantine behavior across rounds. Institutions with AD scores below \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{min}=0.4$$\end{document} are suspended from subsequent aggregation rounds. Ablation study Table 12 reports the contribution of each novel component presented in the proposed model under the combined heterogeneity condition at \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=0.5$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:=0.7$$\end{document}. Removing the complete non-IID-aware aggregation mechanism and substituting it with standard FedAvg resulted in an accuracy drop of 18.45% on IoT-Flock and 17.54% on CICIoT2023 across all ablation configurations. Removing cluster-weighted aggregation alone, while keeping gradient divergence correction, reduces accuracy by 7.89% on IoT-Flock and 7.64% on CICIoT2023. And removing the gradient divergence correction alone, while keeping cluster weighting, brings accuracy down by 6.10% and 6.13%, respectively. Both components together provide full-aggregation contributions of 18.45% and 17.54%. Removing PBFT gradient filtering is bringing accuracy down by 16.00% on IoT-Flock and 15.23% on CICIoT2023 under the combined condition where \:f=4, Byzantine nodes are active. And this is close to 19.49% and 18.64% drops at \:f=4 without filtering. Removing the reputation-based participation policy alone, while keeping geometric median filtering, results in an accuracy drop of 3.49% and 3.00% on IoT-Flock and CICIoT2, respectively 023. For DSAE adaptation components, removing protocol-aware normalization provides an accuracy drop of 4.36% on IoT-Flock and 3.74% on CICIoT2023. And removing the distribution alignment regularizer leads to accuracy drops of 2.90% and 2.53%. Removing the proximal fine-tuning constraint results in accuracy drops of 2.36% and 2.20%. The combined DSAE adaptation contribution is closely 7.26% and 6.74% across the three components for IoT-Flock and CICIoT2023, respectively. Table 12Component ablation under combined heterogeneity condition.Network ConfigurationIoT-Flock Accuracy (%)IoT-Flock F1 (%)CICIoT2023 Accuracy (%)CICIoT2023 F1 (%)Complete Proposed Model90.8387.9486.4783.12w/o Non-IID-Aware Aggregation (FedAvg substituted)72.3868.4768.9364.83w/o Gradient Divergence Correction84.7381.2680.3476.83w/o Cluster-Weighted Aggregation82.9479.4778.8375.34w/o PBFT Gradient Filtering74.8371.3471.2467.83w/o Reputation-Based Participation87.3484.1383.4780.26w/o Protocol-Aware Normalization86.4783.2882.7379.14w/o Distribution Alignment Regularizer87.9384.7483.9480.47w/o Proximal Fine-Tuning Constraint88.4785.3684.2781.03 Table 12Component ablation under combined heterogeneity condition.Network ConfigurationIoT-Flock Accuracy (%)IoT-Flock F1 (%)CICIoT2023 Accuracy (%)CICIoT2023 F1 (%)Complete Proposed Model90.8387.9486.4783.12w/o Non-IID-Aware Aggregation (FedAvg substituted)72.3868.4768.9364.83w/o Gradient Divergence Correction84.7381.2680.3476.83w/o Cluster-Weighted Aggregation82.9479.4778.8375.34w/o PBFT Gradient Filtering74.8371.3471.2467.83w/o Reputation-Based Participation87.3484.1383.4780.26w/o Protocol-Aware Normalization86.4783.2882.7379.14w/o Distribution Alignment Regularizer87.9384.7483.9480.47w/o Proximal Fine-Tuning Constraint88.4785.3684.2781.03 Component ablation under combined heterogeneity condition. Comparison with state-of-the-art Table 13 presents a complete comparison against published IDS and FL. Among methods evaluated on IoT-Flock and CICIoT2023, Sengan et al.,16 establish the centralized upper bound at 96.73% and 93.58% under pooled training without federated constraints. The proposed model achieves 95.42% and 92.17% under IID federated conditions, trails the centralized ceiling by 1.31% and 1.41%, and maintains 93.17% and 89.84% under the combined non-IID condition. Panahi,21 and Jia et al.,22 report centralized accuracies of 94.73%/89.84% and 93.26%/88.67%, without federation, both of which are exceeded by the proposed model’s federated IID performance of 95.42%/92.17%. Table 13Comparison of SOTA.AuthorMethodDatasetTypeIID Acc (%)Non-IID Acc (%)Sengan et al.,16Centralized BC + DLIoT-Flock / CICIoT2023Multiclass96.73 / 93.58N/APanthakkan et al.,20Hybrid DL-IDSIoT-Flock / CICIoT2023Multiclass92.14 / 87.39N/APanahi,21DL-based IDSIoT-Flock / CICIoT2023Multiclass94.73 / 89.84N/AJia et al.,22FlowGuardIoT-Flock / CICIoT2023Multiclass93.26 / 88.67N/AGupta et al.,23BIDSIoT-Flock / CICIoT2023Multiclass89.47 / 85.92N/AFatema,19FL-IDSCICIoT2023Multiclass88.20N/ABegum et al.,17FL-CNN-BiLSTMEdge-IIoTSet / TON-IoTMulticlass97.43 / 98.21N/APeng et al.,25Federated DLEdge-IIoTset / N-BaIoTMulticlass94.82 / 87.7093.86 / 83.81Olanrewaju-George and Pranggono,26FL-IDSN-BaIoTMulticlass90.93N/AWen et al.,27Federated IDSUNSW-NB15Multiclass91.38N/AThis workProposed ModelIoT-Flock/CICIoT2023Multiclass95.42/92.1793.17/89.84N/A — non-IID condition not evaluated. Table 13Comparison of SOTA.AuthorMethodDatasetTypeIID Acc (%)Non-IID Acc (%)Sengan et al.,16Centralized BC + DLIoT-Flock / CICIoT2023Multiclass96.73 / 93.58N/APanthakkan et al.,20Hybrid DL-IDSIoT-Flock / CICIoT2023Multiclass92.14 / 87.39N/APanahi,21DL-based IDSIoT-Flock / CICIoT2023Multiclass94.73 / 89.84N/AJia et al.,22FlowGuardIoT-Flock / CICIoT2023Multiclass93.26 / 88.67N/AGupta et al.,23BIDSIoT-Flock / CICIoT2023Multiclass89.47 / 85.92N/AFatema,19FL-IDSCICIoT2023Multiclass88.20N/ABegum et al.,17FL-CNN-BiLSTMEdge-IIoTSet / TON-IoTMulticlass97.43 / 98.21N/APeng et al.,25Federated DLEdge-IIoTset / N-BaIoTMulticlass94.82 / 87.7093.86 / 83.81Olanrewaju-George and Pranggono,26FL-IDSN-BaIoTMulticlass90.93N/AWen et al.,27Federated IDSUNSW-NB15Multiclass91.38N/AThis workProposed ModelIoT-Flock/CICIoT2023Multiclass95.42/92.1793.17/89.84N/A — non-IID condition not evaluated. Comparison of SOTA. This work Proposed Model N/A — non-IID condition not evaluated. Peng et al.25 constitute the only competing method for reporting non-IID accuracy, achieving 93.86% and 83.81% on the Edge-IIoTset and N-BaIoT datasets. The proposed model achieves 93.17% and 89.84% non-IID accuracies on IoT-Flock and CICIoT2023, under a 3-D combined skew condition with active Byzantine participants. Begum et al.17 report the highest IID federated accuracy of 97.43% and 98.21% on the Edge-IIoTSet and TON-IoT datasets, with different feature sets from those used here, and do not assess under non-IID conditions. Among methods reporting only IID performance, Fatema et al.,19, Olanrewaju-George and Pranggono26, and Wen et al.,27 achieve 88.20%, 90.93%, and 91.38%, respectively, on single datasets without non-IID evaluation, all of which are exceeded by the proposed model’s IID performance on both evaluation datasets. Accuracy, communication, and security trade-off analysis The results across Sect. 7.1–7.6 indicate that the proposed model achieves a favourable trade-off among 3 competing objectives in federated healthcare IoT deployments. From an accuracy perspective, the proposed model achieves 90.83% and 86.47% under combined non-IID conditions on IoT-Flock and CICIoT2023, which are 18.45% and 17.54% higher than FedAvg under the same conditions. From a communication and convergence perspective, the proposed model converges 29 to 67 rounds faster than baselines at moderate to severe heterogeneity levels, as reported in Table 9, which directly reduces total communication rounds and associated transmission cost despite the additional metadata vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{O}(C+2)$$\end{document} per institution per round introduced by the BC transaction wrapper. From a security perspective, PBFT-secured aggregation achieves 87.83% and 83.61% accuracy against four Byzantine users, as reported in Table 11, compared with 71.34% and 67.83% without PBFT filtering — a protection gain of 16.49% and 15.78%. The CL overhead of 234.7 to 467.2 ms reported in Table 10 applies to the aggregation tier only and does not affect the 16 ms edge-level detection latency that meets the clinical real-time constraint. So the proposed model does not require sacrificing any of these 3 objectives to achieve gains in the others-improved non-IID detection accuracy, reduced communication rounds, and Byzantine resilience are achieved together within the same federated model. Conclusion and future work This research work introduced N-IID-AFIDS, which handles statistical heterogeneity and adversarial aggregation corruption in multi-institutional healthcare IoT networks. This model has 4 contributions that differ from those of existing federated IDS. First, a formal non-IID characterization covers quantity, label, and feature skew and provides a basis for testing aggregation mechanisms across controlled levels of heterogeneity. Second, federated adaptations of DSAE, such as protocol-aware normalization and Wasserstein-based distribution alignment regularization, address embedding inconsistency arising from protocol-stack diversity across institutional nodes without requiring raw data exchange. Third, the cluster-weighted aggregation mechanism with gradient divergence correction reduces convergence error due to distributional divergence, achieving 93.17% and 89.84% non-IID detection accuracy on IoT-Flock and CICIoT2023 under combined skew, whereas standard federated averaging achieves 72.38% and 68.93%. Fourth, the PBFT-secured model update validation with geometric median Byzantine filtering and reputation-based participation control maintains 83.61% detection accuracy with 4 adversarial participants, whereas the insecure federation achieves 67.83% under the same conditions. The model also meets the 16 ms clinical real-time detection constraint across both datasets and converges 29–67 rounds faster than other baselines under moderate to severe heterogeneity. This work had several limitations worth noting. Federated evaluation in this study is based on simulated institutional nodes derived from public datasets, rather than from real distributed deployments. The generalization gap of 8.47% between the IoT-Flock and CICIoT2023 transfer directions indicates that cross-domain performance remains an open challenge. A DQN-based MiTM agent requests simulation-based training, which limits adaptation under strict healthcare privacy constraints. The PBFT CL of 467.2 ms at peak load needs more optimization for time-critical clinical alert scenarios. Future work in this area will examine continual learning mechanisms for evolving attack adaptation, homomorphic encryption for zero-knowledge SC execution, and dynamic PBFT sharding for throughput scaling during high-frequency multi-institutional attack surges.