Optimized IoT clustering and assignment in semi-synchronous federated learning Authors: Hadi Farajvand, Nahideh Derakhshanfard, Abbas Mirzaei, Ali Ghaffari, Ali Asghar Pour Haji Kazem Published: 2026 ABSTRACT This study focuses on the important task of optimizing device clustering and assigning them to edge servers, while also implementing data redistribution in hierarchical semi-synchronous federated learning within the realm of advancing edge computing. Our research goal is to increase the performance and scalability of federated learning systems by improving resource allocation and data processing efficiency, which will in turn enhance edge computing frameworks. The current literature does not have thorough methods that can effectively combine model accuracy with optimal device clustering algorithms in hierarchical semi-synchronous federated learning, leading to below-par performance and inefficient use of resources. This difference highlights the need for creative measures that enhance not only model training accuracy but also the grouping of devices as opposed to current methods. The study utilizes a Graph Neural Network (GNN) to group IoT devices according to their hardware features and local datasets, then applies the K-means algorithm to create efficient device clusters. After that, Hybrid Data Redistribution is used to equalize local datasets in each cluster, and Proximal Policy resource allocation optimization algorithm is implemented to allocate devices to edge servers according to bandwidth usage, and energy consumption based on real-time updates, ultimately enabling hierarchical semi-synchronous federated learning to improve model training. The results show a 15% increase in clustering metrics compared to current algorithms, showcasing how our method improves device assignment and data redistribution in hierarchical semi-synchronous federated learning, addressing issues in model accuracy and resource optimization. Introduction THE Federated Learning (FL) is now seen as a cru-cial model for allowing cooperative machine learning among decentralized Internet of Things (IoT) devices while also protecting data privacy1,2. This decentralized method reduces the hazards linked to centralized data gath-ering, making it very appropriate for various IoT settings limited in resources3,4. The importance of FL lies in its capacity to improve model accuracy and scalability through effective device heterogeneity management and resource al-location optimization5,6. Although FL has its benefits, there are various obstacles that hinder its common use in IoT systems. Non-identical and non-independent data distributions among devices can cause lower model performance and convergence problems7,8. Moreover, energy usage and delay are still significant issues, particularly in extensive hierarchical FL arrangements1,6. Previous research has suggested different approaches, like enhanced scheduling algorithms and resource optimiza-tion frameworks, but there are still gaps in achieving the ideal balance between efficiency and accuracy2,9. Driven by these obstacles, this research seeks to enhance FL techniques by tackling the inherent constraints of existing methods. Expanding upon existing hierarchical and clustered FL frameworks1,2, we present a new method for device scheduling and resource allocation that improves scalability and lowers energy usage while maintaining model accu-racy. Moreover, we incorporate sophisticated reinforcement learning methods to dynamically adjust to changing network conditions and data distributions5,10. By addressing the deficiencies highlighted in previous studies, this research aids in the advancement of stronger and more effective FL systems designed specifically for diverse IoT settings. Federated Learning (FL), Edge Computing, and the Inter-net of Things (IoT) are the essential components of modern decentralized intelligence systems. FL allows collaborative training of models on multiple devices without centralizing data, improving data privacy and decreasing communication overhead10. Edge Computing supports Federated Learning11,12 by handling data nearer to its origin, reducing delay and enabling immediate decision-making crucial for IoT applica-tions13. The IoT ecosystem depends on these technologies to provide scalable and efficient solutions14,15. It is described as a large network of diverse devices7,16. FL works based on a decentralized model training ap-proach, where individual devices train local models that are then centrally aggregated to create a global model10,17. Over time, FL has developed to tackle issues like data diversity and communication effectiveness. Hierarchical FL architectures represent a notable progress by adding intermediate aggrega-tion layers to improve scalability and lessen the workload on central servers1,6. These structured systems help stream-line device scheduling and resource management, which are essential for the success of extensive IoT implementations. Edge Computing has a critical role in handling and ana-lyzing data produced by IoT devices. Edge servers manage data clustering and device allocation to maximize resource usage and reduce latency2,8. Current techniques for device clustering frequently depend on algorithms such as K-means and reinforcement learning to categorize devices according to their similarity and resource availability9,18. Nevertheless, these strategies may face challenges in handling scalability and adaptability in changing network environments19. Redistributing data is vital to evenly distributing compu-tational work and ensuring fair use of resources within the network. Present approaches in federated and edge comput-ing environments prioritize enhancing data flow efficiency and reducing redundancy3,20. Even with these attempts, there are still notable deficiencies in combining clustering, device allocation, and data redistribution into a unified sys-tem that can adjust to different network circumstances5,21. Various obstacles continue to exist in the implementation of Federated Learning in Internet of Things (IoT) systems22,23. Combining clustering, device allocation, and data redistribu-tion separately can result in less than optimal performance and inefficiencies5,24. Moreover, expanding hierarchical FL systems to support a growing number of devices without compromising efficiency is still a significant challenge1,6. The management of resources and stability of FL mod-els are made more complex by dynamic and unpredictable network conditions25,26. Disagreements in the industry frequently involve the com-promises between synchronization techniques and the equi-librium of data privacy and model performance. Discussions persist on whether synchronous or semi-synchronous training techniques provide the optimal balance between speed of convergence and utilization of resources27. Furthermore, strict adherence to data privacy measures may impede the ef-fectiveness and precision of FL models, leading to continued debates on the best methods for preserving privacy28. There is a requirement for thorough research in unexplored territories that considers optimal clustering, device alloca-tion, and data redistribution in hierarchical semi-synchronous FL frameworks. Furthermore, a lack of studies implementing these combined techniques in actual IoT situations hinders the usefulness and expansiveness of current solutions29,30. Although hierarchical federated learning (FL) frameworks have improved scalability and efficiency in IoT settings, cur-rent methods frequently view device clustering, assignment, and data redistribution as separate steps5. This division results in less than optimal performance, especially in net-works that are dynamic and heterogeneous, experiencing fluctuations in resource availability and data distributions9,31,32. Imbalanced computational loads and inefficient resource utilization cause higher latency and lower model accuracy. The importance of this issue is found in how it affects the implementation of Federated Learning in IoT systems on a large scale. Poor integration of clustering and assignment mechanisms limits the scalability and responsiveness of FL models due to ineffective adaptation to changing network conditions8,33. Additionally, lacking effective methods for re-distributing data can lead to uneven workloads on devices, in-tensifying energy usage and decreasing system performance overall1. If not dealt with, these restrictions could hinder the ef-fective use of FL in important IoT fields like smart cities and industrial automation, where quick and precise data processing is crucial13. The absence of a unified structure that can adjust to changes in the network hinders the possible advantages of FL, such as improved privacy and decreased communication costs10,34. The main goal of this research is to create a unified frame-work that enhances device grouping, arrangement, and allo-cation in a hierarchical semi-synchronous federated learning (FL) system for diverse IoT settings35. Our goal is to improve efficiency and resource usage by using GNN-Kmeans to cluster devices and assign those clusters to edge servers1,2. Another goal is to apply advanced reinforcement learn-ing methods, such as Proximal Policy Optimization (PPO), for the dynamic scheduling and assigning of devices. This method gives priority to devices according to data diversity tags and resource availability, adjusting to varying network conditions for enhancing model convergence and accuracy5,36,37. Furthermore, our goal is to integrate analysis of di-verse data and labeling into the decision-making process for scheduling, to prioritize devices that provide valuable and varied data in the FL process8,27. Strategies like FedProx and Gradual Mixing will be applied to manage variations in devices and improve convergence, alleviating problems associated with non-IID data distributions29,38. Our hypothesis is that combining optimized device group-ing with GNN-Kmeans, dynamic scheduling and assignment using PPO, and stabilization methods such as FedProx and Gradual Mixing will result in a notable decrease in latency and enhancement in model accuracy within hierarchical semi-synchronous39 FL systems. This integration tackles the scalability, efficiency, and adaptability hurdles in diverse IoT environments. The study aims to improve the effectiveness and expandability of FL in IoT, addressing previous research gaps and advancing decentralized intelligence systems1,5. Methodology This study employs a simulation-based experimental ap-proach to develop and evaluate a comprehensive framework aimed at optimizing device clustering, scheduling, and as-signment within a hierarchical semi-synchronous federated learning system tailored for heterogeneous IoT environ-ments. Hierarchical semi-synchronous federated learning framework The proposed framework is structured in a multi-tiered hi-erarchy comprising IoT devices, edge servers, and a central cloud server. Edge servers manage clusters of devices, co-ordinating local training and aggregation. The central server periodically synchronizes with the edge servers to update the global model. A semi-synchronous mechanism is im-plemented, allowing devices and edge servers to conduct training and updates at their own pace while maintaining sufficient synchronization to ensure model convergence. Clustering and device assignment strategy To effectively manage the diversity and dynamics of IoT de-vices, a Graph Neural Network (GNN)-based K-means clus-tering algorithm is utilized. Devices are represented as nodes within a graph, with edges capturing relationships and simi-larities based on features such as computational capabilities, network latency, data distribution, and energy availability. The GNN processes this graph to generate embeddings for each device, encapsulating complex relationships and feature interactions. These embeddings are subsequently clustered using the K-means algorithm, grouping devices with sim-ilar characteristics and relationships.Clusters are allocated to edge servers by considering factors such as geographical proximity and resource availability, thereby promoting ef-ficient communication and computation within clusters and optimizing resource utilization across the network. Data redistribution and labeling based on diversity To enhance model generalization and address non-IID data distributions, the local data on each device is analyzed for diversity. Diversity metrics, including label distribution entropy and statistical distance measures, are employed to evaluate how well each device’s data represents the overall distribution. Based on these metrics, devices are categorized to inform scheduling decisions.Data redistribution strate-gies are implemented to balance computational loads and improve data diversity within clusters. Techniques such as intra-cluster data sharing, synthetic data generation, and data augmentation are utilized to enhance the diversity of data available for local training. Dynamic device scheduling and reassignment A Proximal Policy Optimization (PPO) reinforcement learn-ing agent is employed as a scheduling agent to dynamically select devices for participation in each training round. The agent takes into account data diversity labels, resource avail-ability, device reliability, and network conditions to make optimal scheduling decisions. This dynamic scheduling pri-oritizes devices that contribute valuable and diverse data, thereby improving model convergence and accuracy.The sys-tem supports dynamic device reassignment, allowing devices to be reallocated to different clusters or edge servers in response to changes in device states, network conditions, or data distributions. This adaptability enhances the system’s resilience and efficiency in the face of environmental fluctuations. Federated learning process Local Training with FedProx: Devices train local models using the FedProx algorithm, which incorpo-rates a proximal term in the loss function to limit divergence from the global model and ensure stable training across heterogeneous devices.Edge Aggregation: Devices transmit their local model updates to their respective edge servers. Edge servers aggregate these updates to form updated edge models using techniques such as weighted averaging based on device data sizes.Global Aggregation: The central cloud server period-ically receives aggregated models from edge servers. The global model is updated by aggregating the edge models on the cloud server.Gradual Integration and Incremental Updating: To maintain a balance between global and local objectives and improve convergence, models are gradually mixed, and shared layers are updated layer by layer to prevent abrupt changes in model parameters. Local Training with FedProx: Devices train local models using the FedProx algorithm, which incorpo-rates a proximal term in the loss function to limit divergence from the global model and ensure stable training across heterogeneous devices. Local Training with FedProx: Devices train local models using the FedProx algorithm, which incorpo-rates a proximal term in the loss function to limit divergence from the global model and ensure stable training across heterogeneous devices. Edge Aggregation: Devices transmit their local model updates to their respective edge servers. Edge servers aggregate these updates to form updated edge models using techniques such as weighted averaging based on device data sizes. Edge Aggregation: Devices transmit their local model updates to their respective edge servers. Edge servers aggregate these updates to form updated edge models using techniques such as weighted averaging based on device data sizes. Global Aggregation: The central cloud server period-ically receives aggregated models from edge servers. The global model is updated by aggregating the edge models on the cloud server. Global Aggregation: The central cloud server period-ically receives aggregated models from edge servers. The global model is updated by aggregating the edge models on the cloud server. Gradual Integration and Incremental Updating: To maintain a balance between global and local objectives and improve convergence, models are gradually mixed, and shared layers are updated layer by layer to prevent abrupt changes in model parameters. Gradual Integration and Incremental Updating: To maintain a balance between global and local objectives and improve convergence, models are gradually mixed, and shared layers are updated layer by layer to prevent abrupt changes in model parameters. Implementation details GNN-Kmeans Clustering: Implemented using Py-Torch Geometric to leverage graph-oriented deep learning features.Federated Learning Simulation: Modeled using Ten-sorFlow Federated, providing flexibility for simulating federated learning scenarios.Scheduling Agent: The PPO agent for device scheduling is implemented using reinforcement learning li-braries, enabling efficient learning within the simulation.Simulation Environment: Devices and edge servers are simulated with varying computational capabilities, network latencies, and energy constraints to reflect real-world IoT environments. GNN-Kmeans Clustering: Implemented using Py-Torch Geometric to leverage graph-oriented deep learning features. GNN-Kmeans Clustering: Implemented using Py-Torch Geometric to leverage graph-oriented deep learning features. Federated Learning Simulation: Modeled using Ten-sorFlow Federated, providing flexibility for simulating federated learning scenarios. Federated Learning Simulation: Modeled using Ten-sorFlow Federated, providing flexibility for simulating federated learning scenarios. Scheduling Agent: The PPO agent for device scheduling is implemented using reinforcement learning li-braries, enabling efficient learning within the simulation. Scheduling Agent: The PPO agent for device scheduling is implemented using reinforcement learning li-braries, enabling efficient learning within the simulation. Simulation Environment: Devices and edge servers are simulated with varying computational capabilities, network latencies, and energy constraints to reflect real-world IoT environments. Simulation Environment: Devices and edge servers are simulated with varying computational capabilities, network latencies, and energy constraints to reflect real-world IoT environments. Evaluation metrics and performance indicators Model Accuracy: Assessed using metrics such as accuracy, precision, recall, and F1-score on a test dataset.Convergence Rate: Measured by the number of communication rounds required to reach a predefined accuracy threshold.Latency: Evaluated by measuring communication and computation times at each hierarchy level.Resource Utilization: Monitored by tracking computational load, energy consumption, and network band-width usage.Scalability: Assessed by varying the number of devices and edge servers to evaluate performance under different scales. Model Accuracy: Assessed using metrics such as accuracy, precision, recall, and F1-score on a test dataset. Model Accuracy: Assessed using metrics such as accuracy, precision, recall, and F1-score on a test dataset. Convergence Rate: Measured by the number of communication rounds required to reach a predefined accuracy threshold. Convergence Rate: Measured by the number of communication rounds required to reach a predefined accuracy threshold. Latency: Evaluated by measuring communication and computation times at each hierarchy level. Latency: Evaluated by measuring communication and computation times at each hierarchy level. Resource Utilization: Monitored by tracking computational load, energy consumption, and network band-width usage. Resource Utilization: Monitored by tracking computational load, energy consumption, and network band-width usage. Scalability: Assessed by varying the number of devices and edge servers to evaluate performance under different scales. Scalability: Assessed by varying the number of devices and edge servers to evaluate performance under different scales. Experimental setup Datasets: Publicly available datasets, such as MNIST and CIFAR-10, are partitioned in a non-IID manner to simulate real-world data distributions.Baseline Comparison: The framework is compared against traditional federated learning methods that do not incorporate clustering or data diversity considerations.Configurations: Different network configurations are simulated by varying device heterogeneity, network la-tencies, and data distributions. Datasets: Publicly available datasets, such as MNIST and CIFAR-10, are partitioned in a non-IID manner to simulate real-world data distributions. Datasets: Publicly available datasets, such as MNIST and CIFAR-10, are partitioned in a non-IID manner to simulate real-world data distributions. Baseline Comparison: The framework is compared against traditional federated learning methods that do not incorporate clustering or data diversity considerations. Baseline Comparison: The framework is compared against traditional federated learning methods that do not incorporate clustering or data diversity considerations. Configurations: Different network configurations are simulated by varying device heterogeneity, network la-tencies, and data distributions. Configurations: Different network configurations are simulated by varying device heterogeneity, network la-tencies, and data distributions. To ensure clarity and reproducibility of our experimental setup, Table 1 summarizes the key hyperparameters and configuration settings used in the proposed method as well as in all baseline approaches. This compact overview provides a unified reference for model, clustering, optimization, and federated learning parameters, enabling a clear understanding of the comparative evaluation. Table 1Key hyperparameters and configurations.ParameterValueParameterValueDatasetsMNIST, FMNIST, CIFAR-10Image Sizes28 × 28 × 1/32 × 32 × 3GNN layers2GNN Embedding Size64GNN learning rate0.001K-Means (k)5K-means iterations100Agglomerative LinkageWardPPO learning rate0.0003PPO γ0.99PPO λ (GAE)0.95PPO Clip ε0.2PPO policy net2 × 128 MLPGAN Learning Rate0.0002GAN batch size64GAN Epochs100FL rounds100Clients per Round10Local epochs5Local Batch Size32OptimizerSGDFL Learning Rate0.01FedProx µ0.001MNIST/FM ModelDenseNetCIFAR-10 modelDLA—— Table 1Key hyperparameters and configurations.ParameterValueParameterValueDatasetsMNIST, FMNIST, CIFAR-10Image Sizes28 × 28 × 1/32 × 32 × 3GNN layers2GNN Embedding Size64GNN learning rate0.001K-Means (k)5K-means iterations100Agglomerative LinkageWardPPO learning rate0.0003PPO γ0.99PPO λ (GAE)0.95PPO Clip ε0.2PPO policy net2 × 128 MLPGAN Learning Rate0.0002GAN batch size64GAN Epochs100FL rounds100Clients per Round10Local epochs5Local Batch Size32OptimizerSGDFL Learning Rate0.01FedProx µ0.001MNIST/FM ModelDenseNetCIFAR-10 modelDLA—— Key hyperparameters and configurations. Experimental procedure Initialization: The global model is initialized and dis-tributed to edge servers.Clustering and Assignment: Devices are clustered using the GNN-Kmeans algorithm and assigned to edge servers.Local Training: Devices perform local training using the FedProx algorithm.Edge Aggregation: Edge servers aggregate local up-dates from devices.Global Aggregation: The central server updates the global model by aggregating edge models.Dynamic Scheduling: The PPO agent schedules de-vices for subsequent training rounds.Reassessment: Device clusters and assignments are updated as necessary based on reassessment. Initialization: The global model is initialized and dis-tributed to edge servers. Initialization: The global model is initialized and dis-tributed to edge servers. Clustering and Assignment: Devices are clustered using the GNN-Kmeans algorithm and assigned to edge servers. Clustering and Assignment: Devices are clustered using the GNN-Kmeans algorithm and assigned to edge servers. Local Training: Devices perform local training using the FedProx algorithm. Local Training: Devices perform local training using the FedProx algorithm. Edge Aggregation: Edge servers aggregate local up-dates from devices. Edge Aggregation: Edge servers aggregate local up-dates from devices. Global Aggregation: The central server updates the global model by aggregating edge models. Global Aggregation: The central server updates the global model by aggregating edge models. Dynamic Scheduling: The PPO agent schedules de-vices for subsequent training rounds. Dynamic Scheduling: The PPO agent schedules de-vices for subsequent training rounds. Reassessment: Device clusters and assignments are updated as necessary based on reassessment. Reassessment: Device clusters and assignments are updated as necessary based on reassessment. The key contributions of this study are summarized as follows: A GNN-enhanced clustering mechanism that extracts graph-structured embeddings of IoT devices and integrates them with K-means to form robust, topology-aware clusters—unlike static or distance-based approaches used in prior work.A PPO-driven dynamic device-to-edge assignment framework that optimizes allocation based on real-time network conditions, resource status, and device heterogeneity, improving efficiency over heuristic-based methods.A hybrid data redistribution strategy combining GAN-based augmentation with intra-cluster balancing to alleviate Non-IID data issues and data sparsity across devices.A unified semi-synchronous hierarchical FL pipeline, integrating FedProx-based local training, periodic global aggregation, and coordinated clustering/assignment/redistribution—addressing the limitations of prior studies that treated these steps independently. A GNN-enhanced clustering mechanism that extracts graph-structured embeddings of IoT devices and integrates them with K-means to form robust, topology-aware clusters—unlike static or distance-based approaches used in prior work. A GNN-enhanced clustering mechanism that extracts graph-structured embeddings of IoT devices and integrates them with K-means to form robust, topology-aware clusters—unlike static or distance-based approaches used in prior work. A PPO-driven dynamic device-to-edge assignment framework that optimizes allocation based on real-time network conditions, resource status, and device heterogeneity, improving efficiency over heuristic-based methods. A PPO-driven dynamic device-to-edge assignment framework that optimizes allocation based on real-time network conditions, resource status, and device heterogeneity, improving efficiency over heuristic-based methods. A hybrid data redistribution strategy combining GAN-based augmentation with intra-cluster balancing to alleviate Non-IID data issues and data sparsity across devices. A hybrid data redistribution strategy combining GAN-based augmentation with intra-cluster balancing to alleviate Non-IID data issues and data sparsity across devices. A unified semi-synchronous hierarchical FL pipeline, integrating FedProx-based local training, periodic global aggregation, and coordinated clustering/assignment/redistribution—addressing the limitations of prior studies that treated these steps independently. A unified semi-synchronous hierarchical FL pipeline, integrating FedProx-based local training, periodic global aggregation, and coordinated clustering/assignment/redistribution—addressing the limitations of prior studies that treated these steps independently. Comprehensive evaluation on MNIST, Fashion-MNIST, and CIFAR-10, demonstrating up to 15% improvement in clustering metrics and higher global accuracy and communication efficiency compared to state-of-the-art baselines. The computational cost of the proposed framework is primarily dominated by the local federated learning process, since each participating device trains a model on its local dataset. Although auxiliary components such as GNN-based embedding and K-Means clustering introduce additional computational overhead, their cost remains comparatively minor and grows more slowly with the number of devices. As a result, the overall scalability of the system is mainly determined by how the local training computation scales with both the number of devices and the size of their local datasets. Accordingly, the overall computational complexity of the proposed framework can be expressed as: where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:N$$\end{document} denotes the number of participating devices, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m$$\end{document} represents the number of local samples per device, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p$$\end{document} corresponds to the number of model parameters. Related work Federated Learning (FL) has attracted tremendous attention as a decentralized machine learning paradigm, particularly suitable for resource-constrained and privacy-sensitive envi-ronments such as the Internet of Things (IoT). Many works have been devoted to various aspects of FL in an effort to improve its efficiency, scalability, security, and adaptability in heterogeneous IoT networks. This section performs an in-depth analysis of the existing body of literature by categoriz-ing the developments into various principal domains: Device Selection and Clustering, Energy Efficiency and Resource Management, Security and Privacy, Split and Hierarchical Learning, and Integration with Emerging Technologies. Device selection and clustering in federated learning Optimal selection and grouping of devices play a crucial role in enhancing Federated Learning (FL) efficiency in diverse Internet of Things (IoT) networks. Zhang and colleagues3 introduced a spectrum allocation system that saves energy while also selecting robust devices to tackle the issues with non-identically distributed data, resulting in quicker conver-gence and improved training speed. In the same way, Xabier Sáez-de-Cámara and colleagues4 presented a clustered FL structure to detect anomalies in massive and diverse IoT networks, improving model stability and decreasing privacy concerns by using device grouping. Fei Dou and colleagues5 expanded FL by incorporating few-shot learning for indoor positioning on devices, support-ing different types of devices while safeguarding privacy. Yiran Shi and colleagues8 created the Quality-aware De-vice Selection (QDS) algorithm to achieve a balance between training efficiency and accuracy in federated edge learning (FEEL), resulting in enhanced training accuracy and de-creased delays. Furthermore, Fangtong Zhou and colleagues21 specifically studied the improvement of scalability in non-IID FL scenarios by selecting coresets and reducing dimensionality for efficient data distribution estimation. These research studies emphasize the significance of care-fully choosing devices and grouping them strategically to reduce data differences and improve FL effectiveness. Nev-ertheless, challenges like expensive computations and depen-dence on precise device profiling continue to be widespread4,8,21. Energy efficiency and resource management Energy efficiency is an important issue when deploying FL on IoT networks. Rami Hamdi et al.6 focused on energy-efficient hierarchical FL in heterogeneous networks through wireless energy transfer (WET) to lower energy usage and enhance training efficiency. In a similar manner, Zhuoer Hu and colleagues8 presented the Clustered-FedProx method for detecting anomalies in Industrial IoT (IIoT), improving energy efficiency and ensuring stable training. Wei Yang and colleagues9 created the cFL method, which uses cosine similarity and edge-based clustering to decrease communication overhead by 60% and enhance convergence time by 10.3%. Furthermore, Zheng Lin and colleagues40 introduced AdaptSFL, an adaptive split fed-erated learning framework that enhances model aggregation and splitting techniques, greatly decreasing training time in edge networks with limited resources. Although these methods successfully improve energy ef-ficiency, scalability in very dynamic environments and real-world validation still require additional research6,9,40 Table 2. Table 2Summary of selected articles.Ref CiteMethodRedistributionArchitecture layersDataset 5 Federated Learning with Few-Shot Learning for On-Device Indoor PositioningNot specifiedFederated Learning integrated with few-shot learningNot specified 9 CFC: Data Distribution Edge Clustering Algorithm Based on Deep Federated LearningCoreset selection and dimension reductionEncoder-based clustering with K-meansFEMNIST and OpenImage datasets 22 AdaptSFL: Adaptive Split Federated Learning in Resource-Constrained Edge NetworksNot specifiedAdaptive model aggregation and splitting strategiesCIFAR-10 and MNIST datasets 14 Efficient Data Distribution Estimation for Accelerated Federated LearningCoreset selection and dimension reductionEncoder-based clustering with K-meansFEMNIST and OpenImage datasets 23 EPSL: Efficient Parallel Split Learning Over Resource-Constrained Wireless Edge NetworksNot specifiedLayer-wise partitioning and gradient aggregationSimulated wireless network data 18 NIDS-FGPA: Federated Learning Network Intrusion Detection with Secure AggregationNot specifiedFederated model aggregation with secure communicationEdge-IIoTset, CIC IoT 2023 datasets 29 Efficient Collaborations through Weight-Driven Coalition Dynamics in Federated Learning SystemsNot specifiedWeight-driven coalition dynamics with similarity-based clusteringSimulated IoT data-Over-the-Air Hierarchical Personalized Federated LearningNot specifiedThree-tier device-edge-cloud hierarchy with beamformingSimulated environments 24 Split Learning Over Wireless Networks: Parallel Design and Resource ManagementNot specifiedCluster-based parallel split learningSimulated non-IID wireless network data 28 FedorAS: Federated Architecture Search under System HeterogeneityNot specifiedSupernet-based federated NAS with tiered device clusteringCIFAR-10, CIFAR-100, Shakespeare, Speech Commands datasets Table 2Summary of selected articles.Ref CiteMethodRedistributionArchitecture layersDataset 5 Federated Learning with Few-Shot Learning for On-Device Indoor PositioningNot specifiedFederated Learning integrated with few-shot learningNot specified 9 CFC: Data Distribution Edge Clustering Algorithm Based on Deep Federated LearningCoreset selection and dimension reductionEncoder-based clustering with K-meansFEMNIST and OpenImage datasets 22 AdaptSFL: Adaptive Split Federated Learning in Resource-Constrained Edge NetworksNot specifiedAdaptive model aggregation and splitting strategiesCIFAR-10 and MNIST datasets 14 Efficient Data Distribution Estimation for Accelerated Federated LearningCoreset selection and dimension reductionEncoder-based clustering with K-meansFEMNIST and OpenImage datasets 23 EPSL: Efficient Parallel Split Learning Over Resource-Constrained Wireless Edge NetworksNot specifiedLayer-wise partitioning and gradient aggregationSimulated wireless network data 18 NIDS-FGPA: Federated Learning Network Intrusion Detection with Secure AggregationNot specifiedFederated model aggregation with secure communicationEdge-IIoTset, CIC IoT 2023 datasets 29 Efficient Collaborations through Weight-Driven Coalition Dynamics in Federated Learning SystemsNot specifiedWeight-driven coalition dynamics with similarity-based clusteringSimulated IoT data-Over-the-Air Hierarchical Personalized Federated LearningNot specifiedThree-tier device-edge-cloud hierarchy with beamformingSimulated environments 24 Split Learning Over Wireless Networks: Parallel Design and Resource ManagementNot specifiedCluster-based parallel split learningSimulated non-IID wireless network data 28 FedorAS: Federated Architecture Search under System HeterogeneityNot specifiedSupernet-based federated NAS with tiered device clusteringCIFAR-10, CIFAR-100, Shakespeare, Speech Commands datasets Summary of selected articles. Split learning and hierarchical learning Split Learning (SL) is an alternative to FL by splitting models across devices and edge servers for better utilization of re-sources. Zheng Lin et al.41 proposed the EPSL framework for parallel split learning, which lowers training latency and communication overhead. Wen Wu et al.37,42 derived the CPSL (Cluster-based Parallel SL) framework, which further reduces training latency in a non-IID environment. Further, Seychi Lin et al.43 suggested SplitLoRA, com-bining SL with Low-Rank Adaptation (LoRA) for the ef-ficient fine-tuning of large language models, showing a re-duced convergence time and resource requirement. However, these techniques usually involve a high amount of compu-tation and poor generalization in many real heterogeneous settings41–43. Integration with emerging technologies FL integration with other emerging technologies, such as blockchain and deep reinforcement learning (DRL), has shown promise in improving FL capabilities. Romano Fan-tacci and Benedetta Picano44 integrated the Device-to-Device (D2D) communication mechanism and incentive mechanism into FL frameworks to improve the convergence time and energy efficiency in 6G networks. Besides, Wei et al.36 used digital twin and DRL to optimize FL in IIoT, which achieved better training efficiency and lower convergence time. Zheng Lin et al.45 introduced blockchain into hierar-chical FL for IoT to enhance robustness and convergence efficiency but suffered from high computational overhead. Aristeidis Karras et al.46 studied Federated NAS (Fedo-rAS) for dealing with system heterogeneity and obtained substantial performance improvement but was burdened by the scalability issue. These integrations enhance FL’s functionality and effi-ciency but require careful consideration of implementation complexity and scalability36,44–46. Gaps and weaknesses in current research Notwithstanding notable progress, various deficiencies and limitations remain evident in the present landscape of feder-ated learning research: Scalability and Empirical Validation: Most of the works depend on simulated settings usually without concrete analyses of real-world deployment and scal-ability4,6,18,41–43,46.Computational Overheads: Improved algorithms of-ten incur significant computational and communication overheads, rendering them inapplicable to resource-constrained IoT devices4,8,9,18,29,43,47.Dynamic and Adversarial Environments: The cur-rent models mostly assume stable network conditions and do not consider adversarial situations enough, such as model poisoning and data manipulation4,18,20.Handling Heterogeneity: Even though the device clustering and selection techniques try to decrease data heterogeneity, the dynamic nature in IoT network char-acteristics poses extra challenges that those solutions cannot handle fully3,4,8,47.Power Consumption: Although many studies high-light the importance of energy efficiency, achieving a balanced tradeoff between power consumption and computational performance remains a challenge6,9,29,40. Scalability and Empirical Validation: Most of the works depend on simulated settings usually without concrete analyses of real-world deployment and scal-ability4,6,18,41–43,46. Scalability and Empirical Validation: Most of the works depend on simulated settings usually without concrete analyses of real-world deployment and scal-ability4,6,18,41–43,46. Computational Overheads: Improved algorithms of-ten incur significant computational and communication overheads, rendering them inapplicable to resource-constrained IoT devices4,8,9,18,29,43,47. Computational Overheads: Improved algorithms of-ten incur significant computational and communication overheads, rendering them inapplicable to resource-constrained IoT devices4,8,9,18,29,43,47. Dynamic and Adversarial Environments: The cur-rent models mostly assume stable network conditions and do not consider adversarial situations enough, such as model poisoning and data manipulation4,18,20. Dynamic and Adversarial Environments: The cur-rent models mostly assume stable network conditions and do not consider adversarial situations enough, such as model poisoning and data manipulation4,18,20. Handling Heterogeneity: Even though the device clustering and selection techniques try to decrease data heterogeneity, the dynamic nature in IoT network char-acteristics poses extra challenges that those solutions cannot handle fully3,4,8,47. Handling Heterogeneity: Even though the device clustering and selection techniques try to decrease data heterogeneity, the dynamic nature in IoT network char-acteristics poses extra challenges that those solutions cannot handle fully3,4,8,47. Power Consumption: Although many studies high-light the importance of energy efficiency, achieving a balanced tradeoff between power consumption and computational performance remains a challenge6,9,29,40. Power Consumption: Although many studies high-light the importance of energy efficiency, achieving a balanced tradeoff between power consumption and computational performance remains a challenge6,9,29,40. System model The architecture to enhance federated learning for IoT set-tings involves sophisticated device clustering and assignment techniques, data reconfiguration, and semi-synchronous fed-erated learning. The aim is to mitigate heterogeneity, re-source limitation, and communication overhead in large IoT networks with emphasis on scalability and reliability. The basic idea involves grouping Internet of Things (IoT) devices based on their computational capacities, energy availability, and network connections. Graph Neural Networks (GNNs), such as Graph Convolutional Networks (GCNs) and Graph Attention Networks (GATs), infer high-dimensional embed-dings that represent the features of devices and their topolog-ical correlations. The embeddings are then fed to a clustering algorithm like KMeans or Agglomerative Clustering, which groups similar devices, hence simplifying resource allocation and reducing complexity. Clusters are typically assigned to edge servers, resulting in a list of suitable edge servers for devices with similar demands. This improves device deployment and load balancing. Dynamic mapping to edge servers also reduces communication overhead and optimizes resource utilization using a Proximal Policy Optimization (PPO) agent learned in a Deep Reinforcement Learning (DRL) framework. The PPO agent learns the optimal policy of mapping device-to-edge-server through a reward func-tion that negatively weights communication overhead and resource wastage but positively weights load balancing and bandwidth utilization efficiency. The back-off mechanism redistributes devices to less congested servers when an edge server is congested. Heterogeneity and sparsity of data in IoT devices are significant challenges for federated learning. In this paper, we explore the use of Generative Adversarial Networks (GANs) for data augmentation. A GAN trained on actual aggregated data generates synthetic samples for devices with small datasets, enhancing diversity and quan-tity. Better local model training results in a stronger global federated model. Semi-synchronization in federated learning balances efficiency and reliability. FedProx is used by all the edge servers to reduce heterogeneity impacts of data when there is localized model training. The edge servers’ locally trained models are sent asynchronously to the global server, which aggregates updates sporadically with Federated Averaging (FedAvg). This reduces idle time in addition to speeding up model convergence. In all, the approach miti-gates key federated learning challenges in IoT settings by reducing communication overhead through device cluster-ing, ensuring optimum resource allocation through dynamic device assignment using DRL-based PPO agents, mitigating data imbalance through data redistribution using GAN-based algorithms, and achieving a trade-off between asynchronous updates and global aggregation at regular intervals. The method has four main steps: device clustering, device assignment, data redistribution, and semi-synchronous federated learning. Device clustering Optimization of resource allocation and communication pro-cesses in federated learning within Internet of Things (IoT) settings is of utmost importance. To tackle these issues, a device clustering methodology based on graph-based feature extraction coupled with advanced clustering algorithms is introduced. The core hypothesis of this research work posits that graph-based feature extraction methodologies combined with either KMeans or Agglomerative Clustering are capa-ble of providing groupings of IoT devices that accurately represent intrinsic characteristics in both computation and network communication. This grouping further optimizes resource allocation and efficiency in communication within the context of a federated learning setting Fig. 1. Fig. 1General architecture. General architecture. Let a set of IoT devices be denoted by D ={d1,d2,…,dn}. Each device di is associated with a feature vector xi, encapsulating attributes such as computational ca-pabilities, energy resources, network connectivity, and other pertinent characteristics. These feature vectors collectively form the set X = {x1,x2,…,xn}. The topology of the network of these devices is denoted as a graph G = (V, E), where V is the set of nodes, including both IoT devices and network switches, and E denotes the set of edges representing network connections. To better represent features for each device, a Graph Neural Network (GNN), such as a Graph Convolutional Network (GCN) or Graph Attention Network (GAT), is exploited. The GNN maps the raw feature vectors into enriched embeddings hi for each device di, which can be formulated as 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}$$\:{\mathbf{h}}_{i}=\mathrm{G}\mathrm{N}\mathrm{N}({\mathbf{x}}_{i},{\hspace{0.25em}\hspace{0.05em}}\mathcal{N}(i\left)\right)$$\end{document} where N(di) represents the neighborhood of node di in the graph G. These embeddings encode both individual attributes of each device and the structural information derived from the network topology. Next, clustering algorithms are applied in such a manner that the devices are grouped into disjoint clusters represented as C = {C1,C2,…,Ck}. The KMeans algorithm is employed by minimizing the within-cluster sum of squares, formulated as 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}$$\:\mathop {\min \sum {\:_{j = 1}^C} \sum {{\:_{{h_i} \in \:{{\cal C}_j}}}} \left\| {{h_i} - \mu {\:_j}} \right\|_2^2}\limits_{\{ {{\cal C}_j}\} _{j = 1}^C}$$\end{document} where µj denotes the centroid for cluster Cj. Conversely, Agglomerative Clustering is utilized, where clusters are successively merged according to a linkage criterion, forming a hierarchical structure of clusters; for example, Ward’s method. The output of this clustering is a certain number of well separated clusters, where devices in the same cluster share computational and network similarities. Their homogeneity encourages better allocation in edge servers to efficiently manage both the computational workload and communication requirements of these devices. Algorithm 1Device clustering workflow Device clustering workflow Device assignment Following the effective implementation of device clustering, the next most important aspect in federated learning for the Internet of Things (IoT) ecosystem is adaptive device-to-edge server allocation. This paper considers efficient device distribution as key to attaining a balance in the workload of servers, minimization of communication overhead, and optimum overall system efficacy. To address this challenge, a Deep Reinforcement Learning (DRL) framework using a Proximal Policy Optimization (PPO) agent has been pro-posed. The main hypothesis carried out in this work is that the incorporation of a PPO agent into a DRL architecture enforces adaptable and efficient distribution of IoT devices to edge servers, which decreases overhead in communications and improves the utilization of resources Fig. 2. Fig. 2Workflow of the proposed federated learning methodology. Workflow of the proposed federated learning methodology. The device assignment problem can be formulated under the following model: State Space (S): Each state s ∈ S comprises the current metrics of a device and the status of all edge servers. More precisely, the state is defined by Eq. (3): State Space (S): Each state s ∈ S comprises the current metrics of a device and the status of all edge servers. More precisely, the state is defined by Eq. (3): State Space (S): Each state s ∈ S comprises the current metrics of a device and the status of all edge servers. More precisely, the state is defined by Eq. (3): s {m1,m2,…,mp,l1,l2,…,lk,d1,d2,…,dk} (3). where m1 to mp represent device-related metrics, including communication overhead, training time, and energy consumption, while l1 to lk denote the current load on each of the k edge servers, and d1 to dk denote the disk usage on each edge server. Action Space (A): The action space is defined as A = {a1,a2,…,ak}, where each single action aj corresponds to the assignment of a device to the j-th edge server.Policy (π): The policy is parameterized by θ, which defines the probability of choosing action a given state s. Action Space (A): The action space is defined as A = {a1,a2,…,ak}, where each single action aj corresponds to the assignment of a device to the j-th edge server. Action Space (A): The action space is defined as A = {a1,a2,…,ak}, where each single action aj corresponds to the assignment of a device to the j-th edge server. Policy (π): The policy is parameterized by θ, which defines the probability of choosing action a given state s. Policy (π): The policy is parameterized by θ, which defines the probability of choosing action a given state s. The device assignment optimization is performed using the Proximal Policy Optimization (PPO) algorithm; hence, it is known to be efficient and stable for a wide class of reinforcement learning tasks. The objective of PPO is to maximize the expected cumulative reward J(θ) defined as 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}$$\:J\left( {\theta \:} \right) = {_{({s_t},{a_t})\sim \:\pi {\:_{\theta {\kern 1pt} }}}}\left[ {\sum {\:_{t = 0}^T} \gamma {\:^t}{r_t}} \right]$$\end{document} where rt represents the reward received at time t and γ is the discount factor that balances immediate and future rewards. The PPO algorithm updates the policy πθ by maximizing the clipped surrogate objective function 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}$$\:{{\cal L}^{CLIP}}\left( {\theta \:} \right) = {_t}\left[ {\min \left( {{r_t}(\theta \:){A_t},clip({r_t}(\theta \:),1 - \in ,1 + \in ){A_t}} \right)} \right]$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{t}\left(\theta\:\right)=\frac{{\pi\:}_{\theta\:}({a}_{t}\mid\:{s}_{t})}{{\pi\:}_{{\theta\:}_{\mathrm{old}}}({a}_{t}\mid\:{s}_{t})}$$\end{document} where At denotes the advantage estimate at time t, showing how good action at is compared to the average action. The term ϵ is a hyperparameter that controls the clipping range, preventing very large policy updates that might disrupt learning. Start with the policy parameters θ, which are iteratively optimized to maximize J(θ) so that the agent can eventually make informed decisions on assigning devices to edge servers. Adding a PPO agent into the DRL system is expected to yield the best policy π∗. The policy provides concrete means for the agent to intelligently make flexible, smart assignments for balancing load on servers, reducing commu-nication cost, and improving resource usage. Therefore, this proposed assignment strategy is expected to enhance the per-formance of federated learning systems in IoT applications, even as the number of devices and edge servers increases. Data redistribution In the process of building an effective federated learning system for the Internet of Things, there is a critical need for the data to have appropriate levels of heterogeneity with regard to size and structure. Having small and unbalanced datasets per individual IoT device can lead to poor local training results, thus affecting the global federated model’s quality Fig. 3. Fig. 3Workflow of the proposed federated learning methodology. Workflow of the proposed federated learning methodology. Algorithm 2Device assignment workflow Device assignment workflow To address the above issues, a novel data redistribution method relying on Generative Adversarial Networks (GANs) for data augmentation is proposed. The key assumption is that GAN-generated augmentations enable IoT devices to keep more informative and diverse data, making local training more efficient and the resulting global federated model more robust. Let D be the real-world data produced by IoT devices. The Generator network is defined as G, parameterized by θG, and maps a stochastic noise vector z from a random noise space into fake data samples. The Discriminator network, de-noted as D, represents a network parameterized by θD. This network estimates, for a given data x, the probability that it is realistic. The Generative Adversarial Network (GAN) is built to solve a minimax optimization problem, defined as (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\underset{{\theta\:}_{G}}{\mathrm{m}\mathrm{i}\mathrm{n}}\underset{{\theta\:}_{D}}{\mathrm{m}\mathrm{a}\mathrm{x}}{\hspace{0.25em}\hspace{0.05em}}{\mathbb{E}}_{x\sim\:{p}_{\mathrm{data}}\left(x\right)}\left[\mathrm{l}\mathrm{o}\mathrm{g}D(x;{\theta\:}_{D})\right]+{\mathbb{E}}_{z\sim\:{p}_{z}\left(z\right)}\left[\mathrm{l}\mathrm{o}\mathrm{g}\left(1-D(G(z;{\theta\:}_{G});{\theta\:}_{D})\right)\right]$$\end{document} Here, pz(z) is the prior noise distribution, usually set as a Gaussian distribution. The model is set up as a game between two players. The Discriminator becomes better at distinguishing between real and artificially produced data, while the Generator fine-tunes the creation of synthetic data that the Discriminator cannot correctly classify as fake. a: Training Process. The training of the GAN consists of alternating updates between the Discriminator and the Generator: Discriminator Update: The discriminator D is updated to maximize the probability of correctly classifying real and fake data samples. This involves increasing D(x) for real samples drawn from D and decreasing it for synthetic samples generated by G.Generator Update: Generator G is updated to mini-mize the generator loss so that the Discriminator D is wrong in correctly classifying synthetic data as fake. It encourages G to produce increasingly realistic data samples to fool D. Discriminator Update: The discriminator D is updated to maximize the probability of correctly classifying real and fake data samples. This involves increasing D(x) for real samples drawn from D and decreasing it for synthetic samples generated by G. Discriminator Update: The discriminator D is updated to maximize the probability of correctly classifying real and fake data samples. This involves increasing D(x) for real samples drawn from D and decreasing it for synthetic samples generated by G. Generator Update: Generator G is updated to mini-mize the generator loss so that the Discriminator D is wrong in correctly classifying synthetic data as fake. It encourages G to produce increasingly realistic data samples to fool D. Generator Update: Generator G is updated to mini-mize the generator loss so that the Discriminator D is wrong in correctly classifying synthetic data as fake. It encourages G to produce increasingly realistic data samples to fool D. Through this adversarial training process, the Generator progressively learns to generate synthetic data that closely resembles the real data distribution D. Once the GAN is sufficiently trained, the Generator G can produce a sub-stantial number of synthetic samples xfake to augment the existing datasets of IoT devices. For devices with scarce real data, specifically for those where |Di| < τ (a predefined threshold), the dataset is augmented as (7) Table 3: Table 3Symbol used in the hypothesis analysis and workflows. Symbol Description NTotal number of data pointsiIndex of a specific data pointa(i)Average distance between data point i and all other points in the same clusterb(i)Minimum average distance between data point i and points in the nearest neighboring clusters(i)Silhouette score for data point iTTotal clustering timet_endTime when clustering completest_startTime when clustering startsKNumber of clustersi, jIndices representing specific clustersS_iAverage intra-cluster distance for cluster iM_ijDistance between centroids of clusters i and jR_ijRatio (S_i + S_j) / M_ij between clusters i and jR_iMaximum R_ij for cluster iDBIDavies-Bouldin IndexnNumber of devicesd_iData size assigned to device iμMean data distribution across devicesσStandard deviation of data distribution across devicesCVCoefficient of Variation of data distributionCTotal communication overheadTTotal number of communication roundsc_i,tCommunication cost for device i at round tT_rTotal redistribution timet_end,redistTime when redistribution completest_start,redistTime when redistribution startsU_j,rResource utilization for device j and resource rū_rAverage resource utilization for resource rLTotal latencyL_transmissionLatency due to data transmissionL_processingLatency due to data processingL_queueingLatency due to data waiting in queuesCV_lCoefficient of Variation for load balanceμ_lMean load across devicesσ_lStandard deviation of load across devicesAGlobal model accuracy (%)CNumber of correct predictionsTTotal number of predictionsT_train,tTraining time for round tt_end,train,tTime when training completes for round tt_start,train,tTime when training starts for round tSAverage stalenessmTotal number of updatess_uStaleness for update ut_current,roundCurrent round timestampt_uTimestamp when update u was generatedSymbolDescriptionDIoT devices setd_iIoT device iXFeature vectors setx_iFeature vector of device iGNetwork graph topologyVNodes in network graphv_iNode for device iAEdges in network graphh_iNode embedding of iGCNGraph Conv. NetworkGATGraph Attention NetkClusters / edge serversC_jCluster jμ_jCluster j centroidHNode embeddings matrixθPPO agent parametersπ(a|s; θ)Policy: state s to action as_tState at time ta_tAction at time tr_tReward at time tγR.L. discount factorA_tAdvantage at tτData augmentation thresholde_jEdge server jMGeneral modelEEdge servers / GAN epochsδStorage per data samplel_jLoad on edge server jd_jDisk usage on jD_iDevice i datasetD_totalAggregated datasetG(z; θ_G)GAN generatorD(x; θ_D)GAN discriminatorθ_GGenerator parametersθ_DDiscriminator parameterszGAN noise vectorX_realReal data distributionX_fakeFake data from GANmSynthetic samples countD_j^augAugmented data for jM_global(t)Global model at tM_e(t)Edge server e model at tαEdge server weightLLoss functionL_CECross-entropy lossμFedProx reg. parameterU_tEdge servers with updates at tTTotal F.L. rounds / timestepsu_tUpdates at round t||M_e(t) - M_global(t)||FedProx objective term Table 3Symbol used in the hypothesis analysis and workflows. Symbol Description NTotal number of data pointsiIndex of a specific data pointa(i)Average distance between data point i and all other points in the same clusterb(i)Minimum average distance between data point i and points in the nearest neighboring clusters(i)Silhouette score for data point iTTotal clustering timet_endTime when clustering completest_startTime when clustering startsKNumber of clustersi, jIndices representing specific clustersS_iAverage intra-cluster distance for cluster iM_ijDistance between centroids of clusters i and jR_ijRatio (S_i + S_j) / M_ij between clusters i and jR_iMaximum R_ij for cluster iDBIDavies-Bouldin IndexnNumber of devicesd_iData size assigned to device iμMean data distribution across devicesσStandard deviation of data distribution across devicesCVCoefficient of Variation of data distributionCTotal communication overheadTTotal number of communication roundsc_i,tCommunication cost for device i at round tT_rTotal redistribution timet_end,redistTime when redistribution completest_start,redistTime when redistribution startsU_j,rResource utilization for device j and resource rū_rAverage resource utilization for resource rLTotal latencyL_transmissionLatency due to data transmissionL_processingLatency due to data processingL_queueingLatency due to data waiting in queuesCV_lCoefficient of Variation for load balanceμ_lMean load across devicesσ_lStandard deviation of load across devicesAGlobal model accuracy (%)CNumber of correct predictionsTTotal number of predictionsT_train,tTraining time for round tt_end,train,tTime when training completes for round tt_start,train,tTime when training starts for round tSAverage stalenessmTotal number of updatess_uStaleness for update ut_current,roundCurrent round timestampt_uTimestamp when update u was generatedSymbolDescriptionDIoT devices setd_iIoT device iXFeature vectors setx_iFeature vector of device iGNetwork graph topologyVNodes in network graphv_iNode for device iAEdges in network graphh_iNode embedding of iGCNGraph Conv. NetworkGATGraph Attention NetkClusters / edge serversC_jCluster jμ_jCluster j centroidHNode embeddings matrixθPPO agent parametersπ(a|s; θ)Policy: state s to action as_tState at time ta_tAction at time tr_tReward at time tγR.L. discount factorA_tAdvantage at tτData augmentation thresholde_jEdge server jMGeneral modelEEdge servers / GAN epochsδStorage per data samplel_jLoad on edge server jd_jDisk usage on jD_iDevice i datasetD_totalAggregated datasetG(z; θ_G)GAN generatorD(x; θ_D)GAN discriminatorθ_GGenerator parametersθ_DDiscriminator parameterszGAN noise vectorX_realReal data distributionX_fakeFake data from GANmSynthetic samples countD_j^augAugmented data for jM_global(t)Global model at tM_e(t)Edge server e model at tαEdge server weightLLoss functionL_CECross-entropy lossμFedProx reg. parameterU_tEdge servers with updates at tTTotal F.L. rounds / timestepsu_tUpdates at round t||M_e(t) - M_global(t)||FedProx objective term Symbol used in the hypothesis analysis and workflows. 7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{i}^{\mathrm{aug}}={\mathcal{D}}_{i}{\hspace{0.25em}\hspace{0.05em}}\cup\:{\hspace{0.25em}\hspace{0.05em}}{\left\{G({z}_{j};{\theta\:}_{G})\right\}}_{j=1}^{M}$$\end{document} where m denotes the number of synthetic samples gener-ated and added to the original dataset Di. We anticipate that the integration of GAN-based data aug-mentation into the federated learning framework will yield augmented datasets that can significantly enhance the local training performance of IoT devices. By increasing both the diversity and volume of available data, local models trained on these augmented datasets are expected to achieve better generalization and convergence properties. Semi-synchronous federated learning This paper optimizes the federated learning framework in IoT environments using a semi synchronous paradigm. This approach balances training efficiency and model consistency by allowing edge servers to update models asynchronously and aggregate the global model periodically. The central hypothesis is that this paradigm accelerates convergence and improves overall model performance. Let k be the number of edge servers and E= {e1,e2,…,ek} denote the set of these edge servers. The global model at aggregation step t is defined as Mglobal(t), whereas the model at edge server e after the local training at step t is denoted by Me(t). The weight or aggregation rate assigned to every edge server in the process of global aggregation is defined by αe. To address data scarcity in edge devices, synthetic samples are generated and incorporated into the local training process. In the revised approach, instead of assigning random labels, a more informed labeling strategy is adopted. Specifically, labels are assigned based on the local data distribution or inferred using the current local model through pseudo-labeling. To ensure reliability, only high-confidence predictions are selected for training, thereby reducing noise and preventing performance degradation. Algorithm 3Device redistribution workflow Device redistribution workflow Each edge server e ∈ E performs local training with the FedProx objective, which is mathematically defined by (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\cal L}\left( {M_i^{\left( t \right)}} \right) = {{\cal L}_{CE}}\left( {M_i^{\left( t \right)},{{\cal D}_i}} \right) + \mu \:\left\| {M_i^{\left( t \right)} - M_{global}^{\left( t \right)}} \right\|_2^2$$\end{document} where: LCE is the cross-entropy loss computed on the local dataset De.µ is the FedProx regularization parameter that reduces the divergence between the local and global models. LCE is the cross-entropy loss computed on the local dataset De. LCE is the cross-entropy loss computed on the local dataset De. µ is the FedProx regularization parameter that reduces the divergence between the local and global models. µ is the FedProx regularization parameter that reduces the divergence between the local and global models. a: Global Aggregation with FedAvg. Global aggregation is achieved by Federated Averaging (Fe-dAvg) in each aggregation step t, where it updates the global model to a weighted average of models at the edge servers using the constraint in (9) and (10):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sum\:_{e=1}^{E}{\alpha\:}_{e}=1$$\end{document}10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathbf{M}}_{\mathrm{global}}^{(t+1)}=\sum\:_{e=1}^{E}{\alpha\:}_{e}{\hspace{0.17em}}{\mathbf{M}}_{e}^{\left(t\right)}$$\end{document} b: Semi-Synchronous Mechanism. The semi-synchronous mechanism operates on two main axes: Asynchronous Updates: Edge servers perform local training and send updates to the global server asyn-chronously, without waiting for the completion of train-ing rounds by their peers.Periodic Aggregation: The global server performs ag-gregations either periodically or after receiving reports from a predefined fraction of all k edge servers. Asynchronous Updates: Edge servers perform local training and send updates to the global server asyn-chronously, without waiting for the completion of train-ing rounds by their peers. Asynchronous Updates: Edge servers perform local training and send updates to the global server asyn-chronously, without waiting for the completion of train-ing rounds by their peers. Periodic Aggregation: The global server performs ag-gregations either periodically or after receiving reports from a predefined fraction of all k edge servers. Periodic Aggregation: The global server performs ag-gregations either periodically or after receiving reports from a predefined fraction of all k edge servers. This approach allows the global server to keep track of up-dates to the global model Mglobal, incorporating contributions from all edge servers and ensuring global model coherence despite the intrinsic asynchrony of the updates. c: Convergence Analysis. In light of update obsolescence and aggregation rate α, the convergence behavior for this semi-synchronous approach is investigated. Under appropriate conditions on α, learning rates, and data distribution, as stated in (11):11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\underset{t\to\:{\infty\:}}{\mathrm{l}\mathrm{i}\mathrm{m}}{\mathbf{M}}_{\mathrm{global}}^{\left(t\right)}={\mathbf{M}}^{*}$$\end{document} the following proposition is established: the global model is expected to converge towards the optimal solution as t approaches infinity, yielding efficient learning even in the context of asynchronous updates. Performance evaluation The methodology was evaluated using Python, leveraging li-braries such as TensorFlow and PyTorch for Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), Generative Adversarial Networks (GANs), and Deep Reinforcement Learning (DRL). Common clustering tasks were performed using Scikit-learn, while OpenAI Gym provided the Proximal Policy Optimization (PPO) envfigureironment. Data processing and visualization were achieved using NumPy, Pandas, Matplotlib, and Seaborn. Experiments were con-ducted on a high-end system equipped with an Intel i7 CPU, an NVIDIA GTX 1660 GPU, and 16GB RAM. For comparison purposes, two representative baseline methods, namely ROFL and BFL, are considered in the experimental evaluation. ROFL follows a resource-oriented federated learning strategy, where client participation and coordination are guided by system-level constraints without adaptive clustering or learning-based assignment. BFL represents a basic federated learning setup with static client selection and synchronous aggregation. To ensure fairness and reproducibility, all baseline methods use the same datasets, model architectures, number of clients, communication rounds, and training configurations as the proposed approach, differing only in their coordination and client management mechanisms. Comparison with baselines A comparison study was conducted against two baseline methods: Resource Optimizing Federated Learning (ROFL): This baseline employs static clustering methods like KMeans and heuristic algorithms such as Genetic Al-gorithms (GA) and Particle Swarm Optimization (PSO) for resource allocation. However, ROFL lacks dynamic clustering capabilities and fails to address data imbal-ance, relying solely on synchronous FedAvg for model aggregation7.Blockchain-Based Federated Learning with Device Clustering (BFL): BFL utilizes label similarity-based clustering and blockchain for secure communications. However, it suffers from poor performance in address-ing data imbalance and resource heterogeneity. Further-more, its synchronous FedAvg approach leads to com-munication delays and inefficient resource utilization48.Algorithm 4Semi-synchronous federated learning workflow Resource Optimizing Federated Learning (ROFL): This baseline employs static clustering methods like KMeans and heuristic algorithms such as Genetic Al-gorithms (GA) and Particle Swarm Optimization (PSO) for resource allocation. However, ROFL lacks dynamic clustering capabilities and fails to address data imbal-ance, relying solely on synchronous FedAvg for model aggregation7. Resource Optimizing Federated Learning (ROFL): This baseline employs static clustering methods like KMeans and heuristic algorithms such as Genetic Al-gorithms (GA) and Particle Swarm Optimization (PSO) for resource allocation. However, ROFL lacks dynamic clustering capabilities and fails to address data imbal-ance, relying solely on synchronous FedAvg for model aggregation7. Blockchain-Based Federated Learning with Device Clustering (BFL): BFL utilizes label similarity-based clustering and blockchain for secure communications. However, it suffers from poor performance in address-ing data imbalance and resource heterogeneity. Further-more, its synchronous FedAvg approach leads to com-munication delays and inefficient resource utilization48. Blockchain-Based Federated Learning with Device Clustering (BFL): BFL utilizes label similarity-based clustering and blockchain for secure communications. However, it suffers from poor performance in address-ing data imbalance and resource heterogeneity. Further-more, its synchronous FedAvg approach leads to com-munication delays and inefficient resource utilization48. Semi-synchronous federated learning workflow Proposed method enhancements The proposed approach extends these works by incorporating the following improvements: Adaptive Device Clustering with GCNs/GATs: Un-like the static clustering methods in ROFL and BFL, the proposed approach uses Graph Convolutional Net-works (GCNs) and Graph Attention Networks (GATs) to enable adaptive and context-aware device clustering. This improves resource allocation and reduces commu-nication overhead7,48.Dynamic Device Assignment with DRL-PPO: A Deep Reinforcement Learning-based PPO agent dy-namically assigns devices to edge servers based on real-time metrics. This significantly enhances resource utilization compared to the static or heuristic methods used in ROFL and BFL7.GAN-Based Data Augmentation: To address data imbalance, GANs generate synthetic data that improve the accuracy of local models. This approach outper-forms the static label balancing technique employed in BFL48.Semi-Synchronous Federated Learning: The pro-posed framework combines FedProx for local train-ing with periodic asynchronous updates and FedAvg for global aggregation. This hybrid approach balances model consistency and convergence speed while avoiding the bottlenecks associated with synchronous up-dates in ROFL and BFL7,48. Adaptive Device Clustering with GCNs/GATs: Un-like the static clustering methods in ROFL and BFL, the proposed approach uses Graph Convolutional Net-works (GCNs) and Graph Attention Networks (GATs) to enable adaptive and context-aware device clustering. This improves resource allocation and reduces commu-nication overhead7,48. Adaptive Device Clustering with GCNs/GATs: Un-like the static clustering methods in ROFL and BFL, the proposed approach uses Graph Convolutional Net-works (GCNs) and Graph Attention Networks (GATs) to enable adaptive and context-aware device clustering. This improves resource allocation and reduces commu-nication overhead7,48. Dynamic Device Assignment with DRL-PPO: A Deep Reinforcement Learning-based PPO agent dy-namically assigns devices to edge servers based on real-time metrics. This significantly enhances resource utilization compared to the static or heuristic methods used in ROFL and BFL7. Dynamic Device Assignment with DRL-PPO: A Deep Reinforcement Learning-based PPO agent dy-namically assigns devices to edge servers based on real-time metrics. This significantly enhances resource utilization compared to the static or heuristic methods used in ROFL and BFL7. GAN-Based Data Augmentation: To address data imbalance, GANs generate synthetic data that improve the accuracy of local models. This approach outper-forms the static label balancing technique employed in BFL48. GAN-Based Data Augmentation: To address data imbalance, GANs generate synthetic data that improve the accuracy of local models. This approach outper-forms the static label balancing technique employed in BFL48. Semi-Synchronous Federated Learning: The pro-posed framework combines FedProx for local train-ing with periodic asynchronous updates and FedAvg for global aggregation. This hybrid approach balances model consistency and convergence speed while avoiding the bottlenecks associated with synchronous up-dates in ROFL and BFL7,48. Semi-Synchronous Federated Learning: The pro-posed framework combines FedProx for local train-ing with periodic asynchronous updates and FedAvg for global aggregation. This hybrid approach balances model consistency and convergence speed while avoiding the bottlenecks associated with synchronous up-dates in ROFL and BFL7,48. These enhancements collectively address the limitations of existing baselines and optimize federated learning performance in terms of resource allocation, communication efficiency, data balance, and convergence speed. Silhouette score 1)Purpose. Measures how similar an object is to its own cluster com-pared to other clusters, indicating the quality of clustering. 2) Mathematical Formulation For each data point I in (14), (15), (16): |Ci|14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:a\left(i\right)=\frac{1}{\mid\:{\mathcal{C}}_{i}\mid\:-1}\sum\:_{\begin{array}{c}j\in\:{\mathcal{C}}_{i}\\\:j\ne\:i\end{array}}d(i,j)$$\end{document}15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:b\left(i\right)=\underset{k\ne\:i}{\mathrm{m}\mathrm{i}\mathrm{n}}\left(\frac{1}{\mid\:{\mathcal{C}}_{k}\mid\:}\sum\:_{j\in\:{\mathcal{C}}_{k}}(di,j))\right)$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:s\left(i\right)=\frac{b\left(i\right)-a\left(i\right)}{\mathrm{max}\left\{a\left(i\right),{\hspace{0.17em}}b\left(i\right)\right\}}$$\end{document} The overall Silhouette Score for the dataset is (17):17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{Silhouette\:Score}=\frac{1}{N}\sum\:_{i=1}^{N}s\left(i\right)$$\end{document} where N is the total number of data points. 3) Interpretation. s(i) ranges from − 1 to + 1.s(i) ≈ + 1: Data point is well-clustered.s(i) ≈ 0: Data point lies between clusters.s(i) ≈ − 1: Data point may have been assigned to the wrong cluster. s(i) ranges from − 1 to + 1. s(i) ranges from − 1 to + 1. s(i) ≈ + 1: Data point is well-clustered. s(i) ≈ + 1: Data point is well-clustered. s(i) ≈ 0: Data point lies between clusters. s(i) ≈ 0: Data point lies between clusters. s(i) ≈ − 1: Data point may have been assigned to the wrong cluster. s(i) ≈ − 1: Data point may have been assigned to the wrong cluster. Clustering time 1) Purpose. Measures the computational time required to perform the clustering algorithm. 2) Mathematical Formulation 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}$$\:T={t}_{\mathrm{end}}-{t}_{\mathrm{start}}$$\end{document} where: T: Total clustering time.tend: Time when clustering completes.tstart: Time when clustering starts. T: Total clustering time. T: Total clustering time. tend: Time when clustering completes. tend: Time when clustering completes. tstart: Time when clustering starts. tstart: Time when clustering starts. 3) Interpretation. Lower T indicates a more efficient clustering process.Higher T suggests potential inefficiencies or computa-tional bottlenecks. Lower T indicates a more efficient clustering process. Lower T indicates a more efficient clustering process. Higher T suggests potential inefficiencies or computa-tional bottlenecks. Higher T suggests potential inefficiencies or computa-tional bottlenecks. i Index of a specific data point. a(i) Average distance between data point i and all other points in the same cluster. b(i) Minimum average distance between data point i and points in the nearest neighboring cluster. s(i) Silhouette score for data point i. Davies-bouldin index (DBI) 1) Purpose. Evaluates the average similarity ratio of each cluster with its most similar cluster, considering intra-cluster and inter-cluster distances. 2) Mathematical Formulation For each cluster I in (19) and (20):19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{ij}=\frac{{S}_{i}+{S}_{j}}{d\left({\boldsymbol{\mu\:}}_{i},{\boldsymbol{\mu\:}}_{j}\right)}$$\end{document}20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{i}=\underset{j\ne\:i}{\mathrm{max}{R}_{ij}}$$\end{document} The Davies-Bouldin Index is 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}$$\:\mathrm{DBI}=\frac{1}{K}\sum\:_{i=1}^{K}{R}_{i}$$\end{document} where: K: Number of clusters.Si: Average intra-cluster distance for cluster i.Mij: Distance between centroids of clusters i and j. K: Number of clusters. K: Number of clusters. Si: Average intra-cluster distance for cluster i. Si: Average intra-cluster distance for cluster i. Mij: Distance between centroids of clusters i and j. Mij: Distance between centroids of clusters i and j. 3) Interpretation. Lower DBI indicates better clustering with well-separated and compact clusters.Higher DBI suggests overlapping or poorly separated clusters. Lower DBI indicates better clustering with well-separated and compact clusters. Lower DBI indicates better clustering with well-separated and compact clusters. Higher DBI suggests overlapping or poorly separated clusters. Higher DBI suggests overlapping or poorly separated clusters. Data distribution fairness 1) Purpose. Assesses how evenly data is distributed across devices or clusters, ensuring no device is disproportionately burdened Table 4. 2) Mathematical Formulation in (22), (23) and (24)22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{l}=\frac{1}{V}\sum\:_{i=1}^{V}{l}_{i}$$\end{document}23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{l}=\sqrt{\frac{1}{V}\sum\:_{i=1}^{V}{\left({l}_{i}-{\mu\:}_{l}\right)}^{2}}$$\end{document}24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{CV}}_{l}=\frac{{\sigma\:}_{l}}{{\mu\:}_{l}}$$\end{document} where: n: Number of devices.di: Data size assigned to device i.µ: Mean data distribution across devices.σ: Standard deviation of data distribution. Table 4Symbol for evaluation section. Symbol Description NTotal number of data pointsiIndex of a specific data pointa(i)Average distance between data point i and all other points in the same clusterb(i)Minimum average distance between data point i and points in the nearest neighboring clusters(i)Silhouette score for data point iTTotal clustering timet_endTime when clustering completest_startTime when clustering startsKNumber of clustersi, jIndices representing specific clustersS_iAverage intra-cluster distance for cluster iM_ijDistance between centroids of clusters i and jR_ijRatio (S_i + S_j) / M_ij between clusters i and jR_iMaximum R_ij for cluster iDBIDavies-Bouldin IndexnNumber of devicesd_iData size assigned to device iμMean data distribution across devicesσStandard deviation of data distribution across devicesCVCoefficient of Variation of data distributionCTotal communication overheadTTotal number of communication roundsc_i,tCommunication cost for device i at round tT_rTotal redistribution timet_end,redistTime when redistribution completest_start,redistTime when redistribution startsU_j,rResource utilization for device j and resource rū_rAverage resource utilization for resource rLTotal latencyL_transmissionLatency due to data transmissionL_processingLatency due to data processingL_queueingLatency due to data waiting in queuesCV_lCoefficient of Variation for load balanceμ_lMean load across devicesσ_lStandard deviation of load across devicesAGlobal model accuracy (%)CNumber of correct predictionsTTotal number of predictionsT_train,tTraining time for round tt_end,train,tTime when training completes for round tt_start,train,tTime when training starts for round tSAverage stalenessmTotal number of updatess_uStaleness for update ut_current,roundCurrent round timestampt_uTimestamp when update u was generated n: Number of devices. n: Number of devices. di: Data size assigned to device i. di: Data size assigned to device i. µ: Mean data distribution across devices. µ: Mean data distribution across devices. σ: Standard deviation of data distribution. σ: Standard deviation of data distribution. Table 4Symbol for evaluation section. Symbol Description NTotal number of data pointsiIndex of a specific data pointa(i)Average distance between data point i and all other points in the same clusterb(i)Minimum average distance between data point i and points in the nearest neighboring clusters(i)Silhouette score for data point iTTotal clustering timet_endTime when clustering completest_startTime when clustering startsKNumber of clustersi, jIndices representing specific clustersS_iAverage intra-cluster distance for cluster iM_ijDistance between centroids of clusters i and jR_ijRatio (S_i + S_j) / M_ij between clusters i and jR_iMaximum R_ij for cluster iDBIDavies-Bouldin IndexnNumber of devicesd_iData size assigned to device iμMean data distribution across devicesσStandard deviation of data distribution across devicesCVCoefficient of Variation of data distributionCTotal communication overheadTTotal number of communication roundsc_i,tCommunication cost for device i at round tT_rTotal redistribution timet_end,redistTime when redistribution completest_start,redistTime when redistribution startsU_j,rResource utilization for device j and resource rū_rAverage resource utilization for resource rLTotal latencyL_transmissionLatency due to data transmissionL_processingLatency due to data processingL_queueingLatency due to data waiting in queuesCV_lCoefficient of Variation for load balanceμ_lMean load across devicesσ_lStandard deviation of load across devicesAGlobal model accuracy (%)CNumber of correct predictionsTTotal number of predictionsT_train,tTraining time for round tt_end,train,tTime when training completes for round tt_start,train,tTime when training starts for round tSAverage stalenessmTotal number of updatess_uStaleness for update ut_current,roundCurrent round timestampt_uTimestamp when update u was generated Symbol for evaluation section. 3) Interpretation. Lower CV indicates more fairness and uniform data distribution.Higher CV signifies greater disparity in data distribu-tion. Lower CV indicates more fairness and uniform data distribution. Lower CV indicates more fairness and uniform data distribution. Higher CV signifies greater disparity in data distribu-tion. Higher CV signifies greater disparity in data distribu-tion. Communication overhead 1) Purpose. Quantifies the additional data transmitted due to the method, such as synchronization messages, model updates, or coordi-nation data. 2) Mathematical Formulation (25)25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:C=\sum\:_{t=1}^{T}\sum\:_{i=1}^{N}{c}_{i,t}$$\end{document} where: C: Total communication overhead.T: Total number of communication rounds.n: Number of devices.ci, t: Communication cost for device i at round t. C: Total communication overhead. C: Total communication overhead. T: Total number of communication rounds. T: Total number of communication rounds. n: Number of devices. n: Number of devices. ci, t: Communication cost for device i at round t. ci, t: Communication cost for device i at round t. 3) Interpretation. Lower C is preferable as it implies less bandwidth usage and faster communication times.Higher C may indicate inefficiencies in communication protocols or excessive data transmission. Lower C is preferable as it implies less bandwidth usage and faster communication times. Lower C is preferable as it implies less bandwidth usage and faster communication times. Higher C may indicate inefficiencies in communication protocols or excessive data transmission. Higher C may indicate inefficiencies in communication protocols or excessive data transmission. Redistribution time 1) Purpose. Measures the time taken to redistribute data or tasks among devices or clusters after the initial assignment. 2) Mathematical Formulation in (26)26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{r}={t}_{\mathrm{end}}^{\mathrm{redistribution}}-{t}_{\mathrm{start}}^{\mathrm{redistribution}}$$\end{document} where: Tr: Total redistribution time.tend_redistribution: Time when redistribution completes.tstart_redistribution: Time when redistribution starts. Tr: Total redistribution time. Tr: Total redistribution time. tend_redistribution: Time when redistribution completes. tend_redistribution: Time when redistribution completes. tstart_redistribution: Time when redistribution starts. tstart_redistribution: Time when redistribution starts. 3) Interpretation. Shorter Tr enhances overall system efficiency, espe-cially in dynamic environments.Longer Tr may lead to delays and reduced responsive-ness in the system. Shorter Tr enhances overall system efficiency, espe-cially in dynamic environments. Shorter Tr enhances overall system efficiency, espe-cially in dynamic environments. Longer Tr may lead to delays and reduced responsive-ness in the system. Longer Tr may lead to delays and reduced responsive-ness in the system. Resource utilization 1) Purpose Evaluates how effectively computational resources (CPU, GPU, memory) are utilized across devices. 2) Mathematical Formulation in (27) and (28)27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{U}_{j,r}=\frac{{\mathrm{UsedResource}}_{j,r}}{{\mathrm{TotalResource}}_{j,r}}$$\end{document}28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{U}_{r}=\frac{1}{J}\sum\:_{j=1}^{J}{U}_{j,r}$$\end{document} where: Uj, r: Resource utilization for device j and resource r.Ur: Average resource utilization for resource r.n: Number of devices. Uj, r: Resource utilization for device j and resource r. Uj, r: Resource utilization for device j and resource r. Ur: Average resource utilization for resource r. Ur: Average resource utilization for resource r. n: Number of devices. n: Number of devices. 3) Interpretation. Higher average utilization indicates efficient use of re-sources.Excessively high values may lead to bottlenecks, over-heating, or reduced performance.Lower values suggest underutilization of available re-sources. Higher average utilization indicates efficient use of re-sources. Higher average utilization indicates efficient use of re-sources. Excessively high values may lead to bottlenecks, over-heating, or reduced performance. Excessively high values may lead to bottlenecks, over-heating, or reduced performance. Lower values suggest underutilization of available re-sources. Lower values suggest underutilization of available re-sources. Latency 1) Purpose. Measures the time delay experienced in the system, such as data transmission or processing delays. 2) Mathematical Formulation in (29)29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:L={L}_{\mathrm{transmission}}+{L}_{\mathrm{processing}}+{L}_{\mathrm{queueing}}$$\end{document} where: L: Total latency.Ltransmission: Latency due to data transmission.Lprocessing: Latency due to data processing.Lqueueing: Latency due to data waiting in queues. L: Total latency. L: Total latency. Ltransmission: Latency due to data transmission. Ltransmission: Latency due to data transmission. Lprocessing: Latency due to data processing. Lprocessing: Latency due to data processing. Lqueueing: Latency due to data waiting in queues. Lqueueing: Latency due to data waiting in queues. 3) Interpretation. Lower latency is desirable for real-time applications and overall system responsiveness.Higher latency can degrade user experience and system performance, especially in time-sensitive tasks. Lower latency is desirable for real-time applications and overall system responsiveness. Lower latency is desirable for real-time applications and overall system responsiveness. Higher latency can degrade user experience and system performance, especially in time-sensitive tasks. Higher latency can degrade user experience and system performance, especially in time-sensitive tasks. Load balance 1) Purpose. Assesses how evenly the computational or data load is dis-tributed among devices or servers. 2) Mathematical Formulation in (30), (31) and (32)30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{l}=\frac{1}{V}\sum\:_{i=1}^{V}{l}_{i}$$\end{document}31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{l}=\sqrt{\frac{1}{V}\sum\:_{i=1}^{V}{\left({l}_{i}-{\mu\:}_{l}\right)}^{2}}$$\end{document}32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{CV}}_{l}=\frac{{\sigma\:}_{l}}{{\mu\:}_{l}}$$\end{document} where: µl: Mean load across devices.σl: Standard deviation of load across devices.CVl: Coefficient of Variation for load balance.n: Number of devices.li: Load assigned to device i. µl: Mean load across devices. µl: Mean load across devices. σl: Standard deviation of load across devices. σl: Standard deviation of load across devices. CVl: Coefficient of Variation for load balance. CVl: Coefficient of Variation for load balance. n: Number of devices. n: Number of devices. li: Load assigned to device i. li: Load assigned to device i. 3) Interpretation. Lower CVl indicates better load balance.Higher CVl suggests some devices are overburdened while others are underutilized. Lower CVl indicates better load balance. Lower CVl indicates better load balance. Higher CVl suggests some devices are overburdened while others are underutilized. Higher CVl suggests some devices are overburdened while others are underutilized. Global model accuracy 1) Purpose. Evaluates the performance of the federated learning model across all devices on a global test dataset. 2) Mathematical Formulation. A = T × 100%. where: A: Global model accuracy percentage.C: Number of correct predictions.T: Total number of predictions. A: Global model accuracy percentage. A: Global model accuracy percentage. C: Number of correct predictions. C: Number of correct predictions. T: Total number of predictions. T: Total number of predictions. 3) Interpretation. Higher accuracy signifies better model performance and generalization across different datasets.Lower accuracy indicates potential issues with model training, data distribution, or algorithm efficacy. Higher accuracy signifies better model performance and generalization across different datasets. Higher accuracy signifies better model performance and generalization across different datasets. Lower accuracy indicates potential issues with model training, data distribution, or algorithm efficacy. Lower accuracy indicates potential issues with model training, data distribution, or algorithm efficacy. Training time per round 1) Purpose. Measures the duration of each training iteration or round in the federated learning process. 2) Mathematical Formulation in Eq. (33)33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{T}_{\mathrm{train},t}={t}_{\mathrm{end},t}^{\mathrm{train}}-{t}_{\mathrm{start},t}^{\mathrm{train}}$$\end{document} where: Ttrain, t: Training time for round t.tend_train, t: Time when training completes for round t.tstart_train, t: Time when training starts for round t. Ttrain, t: Training time for round t. Ttrain, t: Training time for round t. tend_train, t: Time when training completes for round t. tend_train, t: Time when training completes for round t. tstart_train, t: Time when training starts for round t. tstart_train, t: Time when training starts for round t. 3) Interpretation. Shorter training times per round can lead to faster con-vergence.Longer training times may be necessary for more com-plex models but can reduce overall efficiency. Shorter training times per round can lead to faster con-vergence. Shorter training times per round can lead to faster con-vergence. Longer training times may be necessary for more com-plex models but can reduce overall efficiency. Longer training times may be necessary for more com-plex models but can reduce overall efficiency. Staleness of updates 1) Purpose. Quantifies the delay or lag in the model updates contributed by devices, which can affect convergence and accuracy. 2) Mathematical Formulation. For each update u from device j in (34) and (35):34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{u}={t}_{\mathrm{current}}-{t}_{u}$$\end{document}35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:S=\frac{1}{m}\sum\:_{u=1}^{m}{S}_{u}$$\end{document} where: m: Total number of updates.tcurrent_round: Timestamp of the current round.tu: Timestamp when update u was generated. m: Total number of updates. m: Total number of updates. tcurrent_round: Timestamp of the current round. tcurrent_round: Timestamp of the current round. tu: Timestamp when update u was generated. tu: Timestamp when update u was generated. 3) Interpretation. Lower staleness values indicate more timely updates, enhancing model convergence and accuracy.Higher staleness can lead to outdated updates, poten-tially degrading model performance and slowing con-vergence. Lower staleness values indicate more timely updates, enhancing model convergence and accuracy. Lower staleness values indicate more timely updates, enhancing model convergence and accuracy. Higher staleness can lead to outdated updates, poten-tially degrading model performance and slowing con-vergence. Higher staleness can lead to outdated updates, poten-tially degrading model performance and slowing con-vergence. The scalability of the proposed framework is supported by its hierarchical and distributed design. The computational complexity is mainly driven by local model training, which scales as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:O(N\cdot\:D\cdot\:P)$$\end{document} where N is the number of devices, D is the local dataset size, and P is the number of model parameters. In contrast, the GNN embedding process scales with the graph structure and remains manageable due to sparse connectivity, while K-means clustering operates on compact embeddings. Additionally, the PPO-based scheduling is executed at the edge level, significantly reducing its complexity in large-scale deployments. Therefore, the framework can be extended to thousands of devices by leveraging parallelism across edge servers. D. Ablation Study. To analyze the individual contribution of each module, an ablation study is conducted by selectively disabling key components of the proposed framework. Table 5 summarizes the impact of removing each module on accuracy, clustering quality, and latency, clearly highlighting their respective roles in overall performance improvement. Table 5Ablation study.VariantModules enabledAccuracy (%)DBI ↓Latency (ms) ↓Full modelGNN + PPO + GAN + Semi-Sync88.30.83210w/o GNNPPO + GAN + Semi-Sync85.11.02245w/o PPOGNN + GAN + Semi-Sync86.00.95260w/o GANGNN + PPO + Semi-Sync86.70.91235w/o semi-syncGNN + PPO + GAN85.60.98280 Table 5Ablation study.VariantModules enabledAccuracy (%)DBI ↓Latency (ms) ↓Full modelGNN + PPO + GAN + Semi-Sync88.30.83210w/o GNNPPO + GAN + Semi-Sync85.11.02245w/o PPOGNN + GAN + Semi-Sync86.00.95260w/o GANGNN + PPO + Semi-Sync86.70.91235w/o semi-syncGNN + PPO + GAN85.60.98280 Ablation study. Fig. 4Davies-bouldin index for MNIST dataset. Davies-bouldin index for MNIST dataset. As reported in Table 6, the proposed clustering approach consistently achieves favorable clustering quality across all datasets, as indicated by higher Silhouette scores and lower Davies–Bouldin Index (DBI) values. While the clustering time increases for more complex datasets such as CIFAR-10, the results demonstrate an effective balance between clustering accuracy and computational efficiency. Table 6Clustering metrics.DatasetSilhouette↑DBI↓Time (s)↓MNIST0.540.620.9Fashion-MNIST0.470.711.4CIFAR-100.380.832.8 Table 6Clustering metrics.DatasetSilhouette↑DBI↓Time (s)↓MNIST0.540.620.9Fashion-MNIST0.470.711.4CIFAR-100.380.832.8 Clustering metrics. Fig. 5Silhouette score for MNIST dataset. Silhouette score for MNIST dataset. Figure 4: The Davies-Bouldin Index (DBI) evaluates the quality of clusters produced by different clustering tech-niques on the MNIST dataset. Lower DBI values indicate better cluster separation and compactness. MNIST serves as a simple baseline to compare clustering methods, with lower DBI values suggesting more effective cluster creation in basic scenarios. Figure 5: Silhouette Score on MNIST measures how well each method forms clusters. Since MNIST is simple, better scores here establish a baseline for comparing performance on more complex data, providing a foundation to understand baseline clustering performance before advancing to more complex scenarios. Figure 6: The computational cost of clustering MNIST data is used as a reference. This helps understand how clustering techniques scale when applied to more challenging datasets like Fashion-MNIST or CIFAR-10 for long-term behavior forecasting. Fig. 6Clustering time for MNIST dataset. Clustering time for MNIST dataset. Fig. 7Global model accuracy over rounds for MNIST dataset. Global model accuracy over rounds for MNIST dataset. Figure 7: Global accuracy during training rounds for the MNIST dataset using various federated learning methods. MNIST results serve as a baseline to compare more complex datasets, identifying strategies that converge quickly and es-tablishing a foundation for understanding complexity-related effects. Figure 8: The plot depicts a bar graph of computational cost in clustering versus the number of devices on the x-axis for the Fashion-MNIST dataset. Each bar represents one clustering technique; the higher the bar, the greater the complexity and the better scalability and efficiency of techniques with more complex datasets. Figure 9: compares the Davies-Bouldin Index for Fashion-MNIST. This experiment really tests the ability of clustering algorithms to deal with moderate complexity, ensuring that cluster quality remains high as data complexity incrementally rises from MNIST to CIFAR-10. Figure 10: Clustering quality in terms of Silhouette Score for Fashion-MNIST dataset—compares the ability of differ-ent clustering methods to form good clusters. This figure provides a slightly harder test than MNIST, helping to determine which clustering methods still have high-quality clusters as data complexity rises. Fig. 8Clustering time for fashion-MNIST dataset. Clustering time for fashion-MNIST dataset. Fig. 9Davies-Bouldin Index for fashion-MNIST dataset. Davies-Bouldin Index for fashion-MNIST dataset. Fig. 10Silhouette score for fashion-MNIST dataset. Silhouette score for fashion-MNIST dataset. Figure 11: This figure shows how global model accuracy changes over training rounds for different federated learning methods on Fashion-MNIST. Comparing these trends with MNIST and CIFAR-10 scenarios helps determine how strategies scale and whether moderate dataset complexity enables certain methods to excel in convergence speed and accuracy. Fig. 11Global model accuracy over rounds for fashion-MNIST dataset. Global model accuracy over rounds for fashion-MNIST dataset. Fig. 12Silhouette score for CIFAR-10 dataset. Silhouette score for CIFAR-10 dataset. Figure 12: shows the Silhouette Score on CIFAR-10, which tests how well the different clustering techniques actu-ally group devices into coherent and well-separated clusters. Higher scores indicate better-defined clusters. Since CIFAR-10 is a more complex scenario with much more intricate data distributions, this figure helps to determine which methods. retain clustering quality in spite of added complexity. Figure 13: measures the Davies-Bouldin Index for the CIFAR-10 dataset—it assesses how compact clusters are and how well separated they are. The lower the value, the better the clustering quality. Compared with the Silhouette Score, this metric further points out whether some clustering methods consistently produce well-structured groups under the complexity posed by CIFAR-10 data. Figure 14: This figure shows the time taken to cluster the CIFAR-10 dataset. It presents the computational require-ments for different numbers of devices so that one may know how complexity can increase the time of clustering. Some methodologies can grow significantly more, which could be informative in their sustainability and efficiency in large and complex scenarios. Fig. 13Davies-Bouldin Index for CIFAR-10 dataset. Davies-Bouldin Index for CIFAR-10 dataset. Fig. 14Clustering time for CIFAR-10 dataset. Clustering time for CIFAR-10 dataset. Fig. 15Global model accuracy over rounds for CIFAR-10 dataset. Global model accuracy over rounds for CIFAR-10 dataset. Figure 15: Tracks how global model accuracy evolves over training rounds for different federated learning methods on CIFAR-10. Observing these trends helps identify which approaches converge faster and achieve higher performance, while also showing the impact of varying device counts on scalability and resilience. As shown in Table 7, the proposed framework consistently outperforms the baseline methods across all datasets in terms of classification accuracy. The performance gap becomes more pronounced on more complex datasets such as CIFAR-10, highlighting the effectiveness of the proposed approach in handling data heterogeneity and challenging learning scenarios. Table 7 Accuracy performance.DatasetProposedBaseline 1Baseline 2MNIST99.1%98.4%97.8%Fashion-MNIST93.2%91.0%89.7%CIFAR-1088.3%84.1%82.9% Table 7 Accuracy performance.DatasetProposedBaseline 1Baseline 2MNIST99.1%98.4%97.8%Fashion-MNIST93.2%91.0%89.7%CIFAR-1088.3%84.1%82.9% Accuracy performance. Fig. 16Redistribution time for data redistribution strategies. Redistribution time for data redistribution strategies. Fig. 17Communication overhead for device assignment strategies. Communication overhead for device assignment strategies. Figure 16: shows the redistribution time under different strategies as the number of devices increases. Lower redis-tribution times indicate more efficient strategies in which the models can respond quickly to changing conditions without any single device becoming a bottleneck. Figure 17: This figure illustrates the communication over-head for different device assignment strategies with an in-creasing number of devices. Discovering means to reduce the network load is very necessary in large-scale IoT deployment to make it efficient and responsive. Fig. 18Load balance for device assignment strategies. Load balance for device assignment strategies. Figure 18: evaluates the load balancing achieved by dif-ferent device assignment strategies in consideration of a different number of devices. The better the load balancing, the less the resource constraints and the better the overall utilizationthus showing the scalability and fairness of the assignment schemes. As summarized in Table 8, the proposed framework achieves lower communication overhead and improved load balance compared to static and heuristic assignment strategies. In addition, the reduced redistribution time demonstrates the efficiency of the proposed assignment and coordination mechanisms, particularly in moderate-scale device settings. Table 8 Resource & assignment metrics.MetricProposedStaticHeuristicCommunication overhead (MB)105130150Load balance CV0.120.220.28Redistribution time (s, 100 devices)1.021.482.25 Table 8 Resource & assignment metrics.MetricProposedStaticHeuristicCommunication overhead (MB)105130150Load balance CV0.120.220.28Redistribution time (s, 100 devices)1.021.482.25 Resource & assignment metrics. Fig. 19Energy consumption for number of devices. Energy consumption for number of devices. Figure 19 illustrates the energy consumption behavior of the baseline federated learning approach compared to the proposed framework as the number of IoT devices increases. As observed, the baseline method exhibits a rapid growth in energy consumption due to increased communication overhead and inefficient participation of devices. In contrast, the proposed approach demonstrates a significantly slower growth trend. This improvement is primarily attributed to the intelligent device scheduling mechanism based on PPO, which selectively activates devices, as well as the clustering strategy that reduces redundant communication. As a result, the proposed framework achieves better energy efficiency, particularly in large-scale scenarios. These results confirm that incorporating energy-aware decision-making into the learning process can effectively mitigate the energy cost in industrial IoT environments while maintaining system performance. Discussion and limitations Although the proposed framework demonstrates notable improvements in clustering quality, device assignment efficiency, and federated learning performance, several limitations remain that should be acknowledged: Dependence on Simulation-Based Evaluation: The experiments were conducted in a controlled simulation environment. Real-world IoT networks may exhibit more unpredictable device dropouts, unstable wireless conditions, and hardware inconsistencies that could affect clustering, assignment, and convergence behavior.Computational Overhead of GNN-Based Clustering: While GNNs improve clustering quality, they introduce additional computational cost on the edge server. In extremely large-scale networks with tens of thousands of devices, the embedding generation may become a bottleneck without more efficient model compression techniques.Training Cost of PPO-Based Assignment: Reinforcement learning–based assignment requires periodic policy updates. Although PPO is relatively efficient, the training phase may still be costly when network conditions change frequently.GAN-Based Data Redistribution Challenges: Synthesizing realistic samples for complex datasets (e.g., CIFAR-10) remains challenging. GAN instability or mode collapse may lead to lower-quality augmentations, impacting the training of local models.Semi-Synchronous FL Trade-Offs: The semi-synchronous design reduces idle time but may introduce staleness when edge updates vary significantly in timing. Balancing flexibility and convergence stability remains an open challenge.Energy Consumption Considerations: Although energy metrics were evaluated, the current framework does not explicitly optimize power consumption. Energy-aware scheduling would further enhance practicality in battery-powered IoT environments. Dependence on Simulation-Based Evaluation: The experiments were conducted in a controlled simulation environment. Real-world IoT networks may exhibit more unpredictable device dropouts, unstable wireless conditions, and hardware inconsistencies that could affect clustering, assignment, and convergence behavior. Dependence on Simulation-Based Evaluation: The experiments were conducted in a controlled simulation environment. Real-world IoT networks may exhibit more unpredictable device dropouts, unstable wireless conditions, and hardware inconsistencies that could affect clustering, assignment, and convergence behavior. Computational Overhead of GNN-Based Clustering: While GNNs improve clustering quality, they introduce additional computational cost on the edge server. In extremely large-scale networks with tens of thousands of devices, the embedding generation may become a bottleneck without more efficient model compression techniques. Computational Overhead of GNN-Based Clustering: While GNNs improve clustering quality, they introduce additional computational cost on the edge server. In extremely large-scale networks with tens of thousands of devices, the embedding generation may become a bottleneck without more efficient model compression techniques. Training Cost of PPO-Based Assignment: Reinforcement learning–based assignment requires periodic policy updates. Although PPO is relatively efficient, the training phase may still be costly when network conditions change frequently. Training Cost of PPO-Based Assignment: Reinforcement learning–based assignment requires periodic policy updates. Although PPO is relatively efficient, the training phase may still be costly when network conditions change frequently. GAN-Based Data Redistribution Challenges: Synthesizing realistic samples for complex datasets (e.g., CIFAR-10) remains challenging. GAN instability or mode collapse may lead to lower-quality augmentations, impacting the training of local models. GAN-Based Data Redistribution Challenges: Synthesizing realistic samples for complex datasets (e.g., CIFAR-10) remains challenging. GAN instability or mode collapse may lead to lower-quality augmentations, impacting the training of local models. Semi-Synchronous FL Trade-Offs: The semi-synchronous design reduces idle time but may introduce staleness when edge updates vary significantly in timing. Balancing flexibility and convergence stability remains an open challenge. Semi-Synchronous FL Trade-Offs: The semi-synchronous design reduces idle time but may introduce staleness when edge updates vary significantly in timing. Balancing flexibility and convergence stability remains an open challenge. Energy Consumption Considerations: Although energy metrics were evaluated, the current framework does not explicitly optimize power consumption. Energy-aware scheduling would further enhance practicality in battery-powered IoT environments. Energy Consumption Considerations: Although energy metrics were evaluated, the current framework does not explicitly optimize power consumption. Energy-aware scheduling would further enhance practicality in battery-powered IoT environments. These limitations highlight opportunities for future research, including more efficient GNN architectures, adaptive RL-driven assignment strategies, improved generative models, and real-world large-scale deployment studies. Conclusion The main challenges discussed in using federated learning in IoT, such as heterogeneity of devices, sparsity of data, scarcity of resources, and communication overhead, are ad-dressed in this article. An end-to-end system framework for integrating device clustering, data redistribution, dynamic device assignment, and semi-synchronous federated learning is offered to enable efficient and effective federated learning in large-scale heterogeneous IoT systems. The primary benefit of this methodology lies in its struc-tured division into well-defined components. Firstly, Graph Neural Networks (GNNs) receive high-dimensional embed-dings of Internet of Things (IoT) devices to facilitate ef-fective load balancing and resource allocation. Secondly, a Deep Reinforcement Learning (DRL)-inspired Proximal Pol-icy Optimization (PPO) agent provides optimal device-edge resource assignments for reducing communication overhead. Finally, Generative Adversarial Networks (GANs) enhance data augmentation for alleviating issues with class imbalance and data sparsity in IoT environments, resulting in improved training of local models. Semi-synchronous federated learning balances the flexi-bility of asynchronous updates and the stability of periodic global aggregations, trading off convergence rate for robust-ness. It maximizes scalability and efficiency, enabling fed-erated learning for large-scale heterogeneous IoT networks. It is particularly valuable for safety-critical applications in industrial IoT, smart cities, and healthcare, where timeliness and reliability matter. There are some limitations still. Though this framework supports considerations such as heterogeneity, load balanc-ing, and data augmentation, it needs further optimization for high dynamism or high-churning network environments. Energy consumption along with privacy-preserving solutions need to be studied further in a focused manner. Adaptive clus-tering techniques, state-of-the-art privacy-enhancing mecha-nisms, and device failure or data integrity handling need to be investigated by future studies.