Privacy-preserving federated learning for interpretable student at-risk prediction across schools Authors: Mahdee Jodayree, Arman Kavoosi Ghafi, Mostafa Atashafrouz, Mohammad Hossein Shafiabadi Published: 2026 ABSTRACT Early-warning systems for at-risk students increasingly rely on predictive models trained on sensitive educational records. However, centralized learning pipelines raise concerns about privacy, institutional data sovereignty, and auditability, particularly when student-level data are shared across institutions. This study presents Federated Learning for At-Risk Student Prediction with Differential Privacy and Proof-Before-Train (FL-AtRisk-DP-PBT), a federated learning framework for multi-school at-risk prediction that integrates Federated Averaging (FedAvg)-based training, client-side DP, and a PBT protocol for verifiable logging of client participation and model states. The framework uses a single interpretable global logistic-regression classifier and is evaluated under centralized, standard federated, and FL + DP+PBT regimes on three educational datasets: a primary merged cohort of 14,003 students partitioned into 10 simulated schools, the xAPI-Edu-Data click-stream corpus, and the Students Performance in Exams dataset. On the primary dataset, the centralized model achieves 99.14% accuracy, F1 = 0.9915, and area under the curve (AUC) = 0.9998, while the FedAvg and FL + DP+PBT variants achieve 98.61%/0.9863/0.9993 and 98.00%/0.9802/0.9992, respectively. On xAPI and Exams, FL + DP+PBT reaches approximately 93–94% accuracy, F1 ≈ 0.92–0.93, and AUC ≈ 0.97–0.98. Coefficient-based feature-importance analysis indicates that FL + DP+PBT preserves broadly similar interpretation patterns to the centralized and non-private federated baselines. The PBT ablation introduces only small metric changes relative to DP-only federated training. Overall, the results suggest that interpretable federated at-risk prediction can retain competitive utility while keeping student records local and adding privacy-preserving and verifiable training mechanisms. These findings should be interpreted within the evaluated datasets, simulated school partitions, and label definitions.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-026-56609-z. Introduction Universities worldwide are still grappling with the problems of student failure and dropout, particularly in large first-year and online courses, where early identification of struggling learners is necessary for the design and timely delivery of intervention measures. Recent efforts in the fields of learning analytics and educational data mining (EDM) have demonstrated that prediction models trained on Learning Management System (LMS) traces, assessment data, and behavioral signatures can estimate academic risk with high accuracy several weeks before final grades are known1–4. To illustrate, explainable EDM pipelines developed for programming education and blended courses have shown that relatively small feature sets summarizing engagement, assessment, and time-management patterns suffice to power early-warning systems when combined with machine-learning classifiers and post-hoc interpretability methods1–3. At the same time, multi-dimensional prediction frameworks that aggregate demographic, behavioral, and performance indicators are becoming commonplace for supporting fine-grained academic counselling and personalized support4,5. Parallel lines of inquiry have also established connections between prediction and intervention: when risk estimates are fed back to instructors and students through well-designed dashboards and feedback channels, course completion rates and final grades can be improved in a measurable way5–7. Notwithstanding this progress, several desiderata remain unmet in the use of early-warning models at the institutional and cross-institution levels. On the one hand, many recent systems are trained on data from a single course or institution and are therefore susceptible to cohort shift, concept drift, and changes in instructional design4,8. Research efforts that explicitly evaluate temporal drift in prediction performance show that models trained on historical cohorts can experience a sharp drop in accuracy when the course structure, assessments, or student population change, unless they are periodically recalibrated or updated8. On the other hand, a number of high-performing methods rely on complex ensembles or deep architectures whose decision logic is opaque to instructors and students, even when explainable-AI toolkits are deployed ex post1,3. This motivates the use of intrinsically interpretable models whose parameters can be inspected directly, rather than relying only on post-hoc explanations of complex black-box models. This lack of transparency makes it difficult to translate model outputs into actionable pedagogical decisions, and raises concerns about the robustness and stability of explanations across cohorts and models9. Moreover, most existing interventions are integrated in centralized analytics infrastructures, which assume unfettered access to fine-grained student data and offer limited assurances with respect to privacy, data minimization, and compliance with local regulations. Recent systematic reviews on privacy and inclusiveness in learning analytics have pinpointed the tension between data-hungry predictive models and the expectations of confidentiality, consent, and control that students have over their data10,11. These issues are compounded by the widespread adoption of machine learning in education more generally. Exhaustive surveys of machine-learning applications in education document not only the promise of predictive models to address retention and personalization, but also the ethical and governance challenges involved in algorithms that act on sensitive student records at scale12. Recent work on multimodal and immersive learning analytics further enlarges the set of data modalities under consideration, including clickstreams, grades, eye-tracking, audio, and spatial interaction traces, which introduces new concerns regarding surveillance, data protection, and the interpretability of complex, high-dimensional representations13–15. While these findings highlight the richness of the design space for predictive and prescriptive analytics, they also expose a clear practical and empirical gap: it is currently not well understood how simple, transparent, coefficient-based models such as logistic regression perform when trained in privacy-preserving and distributed settings that explicitly respect institutional boundaries and governance constraints. A second, related gap concerns the adaptation of federated learning (FL) and formal privacy mechanisms to early-warning pipelines. While there is a growing interest in the deployment of federated architectures in education, most existing systems are either concerned with more technical questions such as communication efficiency or personalization, or are evaluated on a single dataset without systematic comparison to centralized baselines4,12. Differential Privacy (DP), for its part, has been explored primarily in centralized settings where noise is added to model updates or to aggregate statistics computed on a pooled dataset10. In the student-success prediction literature, fewer studies jointly examine cross-institution training without raw-data sharing, client-side DP with explicit privacy accounting, and coefficient-based interpretability within a single early-warning pipeline. Even fewer works explore how such models behave across heterogeneous datasets and institutional risk profiles, or how closely federated and privacy-preserving variants can approximate the performance of an idealized centralized reference model5–8. In this paper, we take a step towards addressing these gaps by introducing FL-AtRisk-DP-PBT, a Federated Learning for At-Risk student prediction with DP and Proof-Before-Train framework. We instantiate the framework with a deliberately simple and interpretable model—logistic regression—and evaluate it under three training regimes: (1) centralized training on pooled data, (2) federated training using Federated Averaging (FedAvg) without privacy noise, and (3) federated training with client-side differentially private stochastic gradient descent (DP-SGD), equipped with a lightweight PBT mechanism. Using three complementary datasets that capture different educational contexts and feature spaces, we ask to what extent the global at-risk classifier learned under federated and DP-federated training can approximate the predictive performance of the centralized model, and whether the resulting coefficients are stable enough to support meaningful interpretation of risk factors across settings. The choice of logistic regression is intentional: the model is computationally lightweight, communication-efficient in federated training, directly auditable through its coefficients, and more transparent than deep or highly specialized architectures. Our approach therefore does not aim to introduce a new deep-learning architecture; rather, it evaluates whether a transparent linear model can provide a useful and auditable baseline for privacy-preserving, distributed early-warning systems. The contribution of this work is primarily integrative and empirical. We evaluate whether a transparent logistic-regression model can support binary at-risk prediction across centralized, federated, and differentially private federated training regimes while keeping the model specification, feature processing, optimization protocol, and evaluation metrics consistent. The proposed framework combines FedAvg-based cross-school training, client-side DP-SGD, PBT audit logging, and coefficient-based interpretation within a single experimental pipeline. Empirically, we assess this framework on three educational datasets with different feature spaces and risk-label definitions, reporting predictive performance, ROC-based analyses, local-only baselines, repeated-seed stability, membership-inference attack results, and coefficient-based feature-importance agreement across training regimes. In doing so, the study provides evidence on the extent to which privacy-preserving and auditable federated training can retain competitive utility and interpretable decision patterns relative to centralized logistic-regression baselines, without claiming to introduce a new model architecture or a fundamentally new optimization algorithm. The remainder of the paper is structured as follows. Section 2 reviews related work on early-warning systems, FL in education, privacy-preserving analytics, and explainable student-performance prediction. Section 3 describes the datasets, preprocessing pipeline, and modelling framework, including the centralized and federated training procedures, differential-privacy mechanism, and PBT logging. Section 4 reports the experimental setup and results, including predictive performance, ROC analyses, and coefficient-based interpretability across datasets. Section 5 provides a detailed discussion of the findings, including cross-dataset generalization, practical deployment considerations, and limitations. Section 6 concludes with a summary of contributions, implications for policy and practice, and directions for future research on privacy-aware, interpretable FL in educational settings. Related work Learning analytics and educational data mining have long been used to identify students who are likely to fail, drop out, or require additional academic support. Early studies mainly relied on centralized institutional datasets, such as LMS traces, assessment records, demographic variables, and course-engagement indicators. Over time, this literature has expanded from simple statistical classifiers toward more complex machine-learning models, early-warning dashboards, privacy-aware analytics, and explainable prediction systems. In this section, we organize the related work around four themes that are most relevant to the present study: centralized at-risk prediction, FL in education, privacy-preserving learning analytics, and interpretability or explainability in student-performance prediction. At-risk student prediction and early-warning systems A substantial body of work has shown that student failure, dropout, and low academic performance can be predicted from relatively early educational traces. Macfadyen and Dawson16 provided one of the early demonstrations that LMS activity data can support an early-warning system, using logistic regression to identify students likely to fail a course. Their findings are important because they show that even relatively simple and interpretable models can provide actionable risk signals. Similarly, Hu et al.17 used logistic regression and decision trees on clickstream and assessment data from online learning environments, showing that prediction models updated during the semester can provide timely alerts before final outcomes are known. Akçapınar et al.18 further demonstrated the feasibility of early detection by engineering LMS-based features and applying machine-learning classifiers to identify unsuccessful students within the first weeks of a course. Subsequent work has moved from proof-of-concept prediction toward deployment-oriented early-warning systems. Bañeres et al.19 combined predictive modelling, rule-based risk categories, and instructor-facing dashboards for online higher education, showing that early-warning tools can be integrated into institutional workflows. Jayaprakash et al.20 introduced the Open Academic Analytics Initiative, demonstrating that predictive alerts can be linked to interventions and improvements in course completion and grades. Other studies have emphasized task-specific prediction contexts, such as dropout prediction21, introductory programming failure22, cross-course prediction portability23, and institutional formative assessment24. Across these studies, reported performance is often high, and systematic evidence suggests that educational data-mining models commonly achieve strong accuracy for retention, dropout, and failure prediction tasks25. Despite these advances, most early-warning systems still assume centralized access to student data. This assumption limits their applicability in multi-institutional settings, where data governance, consent, local policy, and institutional autonomy may prevent the pooling of raw student records. In addition, high-performing centralized models may rely on rich institutional features that are difficult to share across schools or interpret consistently across cohorts. These limitations motivate the need for early-warning frameworks that can support cross-institution collaboration while keeping student records local. Federated learning in education FL has recently emerged as a promising approach for educational analytics because it allows multiple institutions to train a shared model without directly exchanging raw data. Fachola et al.26 introduced a FL framework for educational data analytics and showed that federated models can achieve performance close to centralized baselines while reducing the need for data transfer. This work provides an important foundation for applying FL to student-success prediction, but its focus is mainly on distributed training and communication efficiency rather than on formal privacy guarantees or interpretable model behavior. More recent FL studies in education have explored different modelling tasks and system designs. Some approaches use FL for grade prediction, course-outcome mapping, dropout prediction, or learning-outcome estimation across campuses or platforms27–32. For example, FLPADPM uses deep temporal models for dropout prediction in MOOC settings27, while FecMap addresses learning-outcome prediction with multi-layer privacy protection31. Other frameworks incorporate FL into Internet-of-Education or LMS-based environments, including CO–PO mapping and emotion-detection applications32,33. These studies demonstrate the growing relevance of FL for educational prediction and monitoring. However, several gaps remain. Many federated education models rely on complex neural or ensemble architectures, which may improve accuracy but reduce transparency for educators and administrators. Others focus on a single course, institution, or task, making it difficult to assess generalization across heterogeneous educational datasets. In addition, formal DP, empirical privacy-attack evaluation, and verifiable participation logging are not consistently integrated into existing federated educational pipelines. The present study addresses this gap by using a simple global logistic-regression model trained under centralized, FedAvg, and FL + DP+PBT regimes across three different datasets. Privacy-preserving learning analytics and DP Privacy is a central concern in learning analytics because educational records often contain sensitive information about performance, behavior, background, engagement, and learning difficulties. Gürsoy et al.34 surveyed privacy-preserving learning-analytics techniques, including DP, secure multiparty computation, and anonymization, and highlighted the trade-off between analytical utility and disclosure risk. Complementing this technical perspective, Mutimukwe et al.35 examined students’ privacy concerns in learning analytics and showed that trust, perceived benefit, and perceived risk influence students’ willingness to share data. These findings underline that privacy-preserving educational analytics is not only a technical problem but also a governance and trust issue. Several studies have proposed privacy-preserving alternatives to direct centralized learning. Liu et al.36 explored synthetic-data generation as a way to support learning analytics while reducing re-identification risks. Zhao et al.37 proposed DP-TabNet, integrating DP with an interpretable deep tabular model for grade prediction. Zhan et al.38 studied privacy–utility preservation in learning analytics and showed that carefully calibrated privacy mechanisms can maintain useful model performance while providing stronger privacy protection. These studies are directly relevant to privacy-aware educational prediction, but most of them operate in centralized settings or do not explicitly address cross-institution heterogeneity. In federated settings, privacy requires additional care. Keeping raw data local does not by itself eliminate privacy risks, because model updates and output probabilities can still leak information about individual records. This motivates the use of client-side DP and empirical privacy evaluation, such as membership-inference attacks. The present work therefore combines FedAvg with client-side DP-SGD and evaluates privacy leakage empirically through membership-inference attack experiments. The PBT mechanism further adds an auditability layer by logging client participation and model-state commitments before local training, without replacing the role of DP in protecting individual student records. Interpretability, fairness, and explanation stability Interpretability is another major requirement for educational prediction systems, particularly when model outputs may influence student-support decisions. While complex models such as ensembles and deep networks can achieve strong predictive performance, their decision logic is often difficult for instructors, administrators, and students to inspect. This has motivated increasing interest in explainable learning analytics and fairness-aware student-performance prediction. Kesgin et al.39 proposed a fairness-aware student-performance prediction framework that combines predictive modelling with group-level fairness analysis and SHAP-based explanations. Tiukhova et al.40 examined the stability of explainable learning-analytics models and showed that feature attributions can vary across models and data splits, raising concerns about the robustness of explanations. Mai et al.41 proposed an explainable belief-rule-base model that combines expert knowledge with data-driven prediction, while Husayn et al.42 introduced an optimized ensemble for interpretable and fair student-performance prediction. These studies show that interpretability and fairness are increasingly recognized as essential components of educational analytics. However, explanation is often treated as a post-hoc layer applied to centralized or complex models. This creates two limitations. First, explanations may be unstable across training regimes, cohorts, or model classes. Second, post-hoc interpretability does not necessarily make the underlying model auditable or easy to deploy in privacy-constrained institutional settings. By contrast, the present work uses logistic regression as an intrinsically interpretable global model. This allows feature importance to be derived directly from standardized coefficients and compared across centralized, federated, and FL + DP+PBT training regimes. The coefficient-based analysis therefore evaluates not only predictive performance but also whether the proposed privacy-preserving federated model preserves broadly similar interpretation patterns to the centralized and non-private federated baselines. Positioning of the present work Collectively, the reviewed literature shows that high predictive performance for at-risk and dropout detection is already achievable, and that complementary concerns such as federated training, privacy preservation, interpretability, and fairness have each been addressed to some extent. However, these dimensions are often studied separately. Centralized early-warning models typically do not address cross-institution data locality; federated education models do not consistently integrate formal DP or empirical privacy-attack evaluation; privacy-preserving learning-analytics methods often remain centralized; and explainable or fairness-aware methods typically do not operate in distributed and privacy-constrained settings. The present study is positioned at the intersection of these strands. It does not claim to introduce a new deep-learning architecture or a fundamentally new optimization algorithm. Instead, its contribution is integrative and empirical. The proposed framework brings together cross-institution federated training, client-side DP, and a globally interpretable logistic-regression model, augmented with a PBT mechanism for added accountability. This design allows us to examine predictive utility, privacy leakage, auditability, and interpretability within a single experimental framework across three heterogeneous educational datasets. In this way, our work directly responds to the gaps identified in the related literature, offering a concrete proof of concept that competitive early-warning performance can be maintained while adding privacy-preserving training, model transparency, and participation auditability under a unified empirical framework. Proposed method In this section, we describe the proposed method: a privacy-preserving FL framework for predicting at-risk students across multiple schools. The framework consists of three main components – data modelling and multi-school scenario construction, a FL algorithm with DP, and a PBT protocol for verifying client participation – designed to address the practical limitations of educational data: legal/ethical restrictions on centralizing educational records, heterogeneous data distributions across schools, and requirements for both formal privacy guarantees and auditability. Figure 1 illustrates the high-level workflow of the proposed framework, from local preprocessing and at-risk label generation on each client to decentralized training, secure aggregation and verification, which also motivates the subsection structure for the rest of this section. Fig. 1Schematic of the proposed FL-AtRisk-DP-PBT framework with data modelling and multi-school simulation, the unified prediction model under centralized and federated (privacy-preserving with DP) training and the PBT protocol for verifiable client participation. Schematic of the proposed FL-AtRisk-DP-PBT framework with data modelling and multi-school simulation, the unified prediction model under centralized and federated (privacy-preserving with DP) training and the PBT protocol for verifiable client participation. Overall architecture of the proposed federated framework The proposed federated framework adopts a centralized server-multiple clients architecture, where each client corresponds to a distinct school holding its own student records. Let there be K schools, and denote by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}={\left\{\left({x}_{i}^{\left(k\right)},{y}_{i}^{\left(k\right)}\right)\right\}}_{i=1}^{{n}_{k}}$$\end{document} the local dataset of school k, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}^{\left(k\right)}$$\end{document} the feature vector capturing demographic, behavioral, and performance-related attributes and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}^{\left(k\right)}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} the at-risk label. The central goal of the framework is to learn a global prediction model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{w}(\cdot\:)$$\end{document}, parameterized by w, that can accurately estimate the probability that a student is at risk, while ensuring that raw data never leave the respective schools. This setting reflects realistic institutional constraints in education, where data owners (schools) retain control over their records, and only abstract model information is exchanged. Formally, the FL problem is cast as minimizing a weighted sum of local empirical risks across schools. For each school k, the local objective is:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{k}\left(w\right)=\frac{1}{{n}_{k}}\sum\:_{i=1}^{{n}_{k}}\:\mathcal{l}\left({f}_{w}\left({x}_{i}^{\left(k\right)}\right),{y}_{i}^{\left(k\right)}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{l}(\cdot\:,\cdot\:)$$\end{document} is an appropriate loss function (e.g., binary cross-entropy for at-risk classification). The global objective optimized by the framework is:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\underset{w}{\mathrm{m}\mathrm{i}\mathrm{n}}\:F\left(w\right)\triangleq\:\underset{w}{\mathrm{m}\mathrm{i}\mathrm{n}}\:\sum\:_{k=1}^{K}\:\frac{{n}_{k}}{N}{F}_{k}\left(w\right),\:N=\sum\:_{k=1}^{K}\:{n}_{k}$$\end{document} Optimization proceeds in communication rounds: at round \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{t}_{r}$$\end{document} the server distributes the current global parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}^{\left(t\right)}$$\end{document} to a subset or all schools; each selected school performs several steps of local training on \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}$$\end{document} to obtain updated parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{k}^{(t+1)}$$\end{document}; the server then aggregates the updates into a new global model, for example using a FedAvg rule:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}^{(t+1)}=\sum\:_{k\in\:{\mathcal{S}}^{\left(t\right)}}\:\frac{{n}_{k}}{\sum\:_{j\in\:{\mathcal{S}}^{\left(t\right)}}\:\:{n}_{j}}{w}_{k}^{(t+1)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{S}}^{\left(t\right)}$$\end{document}denotes the set of participating schools at round t. This architecture naturally accommodates heterogeneous (non-IID) data distributions and varying client sizes. On top of this basic FL architecture, the framework integrates two additional layers: DP and a PBT protocol. DP is enforced at the client side by perturbing local updates before transmission, typically via gradient or parameter clipping followed by the addition of calibrated random noise; abstractly, each school applies a randomized mechanism \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{M}}_{k}$$\end{document} to its update so that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{M}}_{k}$$\end{document} satisfies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:(\epsilon\:,\delta\:)$$\end{document}-DP with respect to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}$$\end{document}. In parallel, the PBT protocol augments each communication round with cryptographic commitments (and optionally signatures) sent from clients to the server before local training, binding each client to a specific round identifier and received model state. The server verifies these proofs when receiving the noisy updates, thereby obtaining a verifiable log of which schools participated in which rounds and under which model state, without accessing any raw data or undoing the DP guarantees. Collectively, this architecture provides a coherent end-to-end design for privacy-preserving, verifiable at-risk student prediction across multiple schools. Data modelling and multi-school scenario construction This subsection describes how raw records from the educational datasets are mapped to a unified at-risk prediction task and how a realistic multi-school federated scenario is constructed from them. From the raw student records, we model each instance as a feature vector representing the demographic, behavioral, and performance-related attributes, and a binary at-risk label determined by the corresponding academic outcome (e.g., grades/exam scores). The centralized dataset obtained in this way is split into a number of disjoint subsets, where each of them corresponds to a client, and contains data with a local distribution, size, and at-risk ratio, simulating a school and thus inducing a heterogeneous (non-IID) multi-school scenario. The same data modelling and partitioning process is applied to the main dataset used for the development and tuning of the overall framework in the first place, and then identically replicated on the other two datasets. This, in turn, provides the basis for the detailed descriptions of the FL, DP, and PBT in the following two subsections. At-risk label definition and prediction targets The at-risk label in this study is defined as a binary outcome variable that encodes whether a student is considered academically at risk according to observable performance indicators. Formally, for each student i, let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} denote the target label, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}=1$$\end{document} indicates that the student is at risk and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}=0$$\end{document} otherwise. The label is obtained by applying a dataset-specific mapping \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}^{\left(m\right)}(\cdot\:)$$\end{document} to the raw performance related attributes in dataset m, such that4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}^{\left(m\right)}={g}^{\left(m\right)}\left({r}_{i}^{\left(m\right)}\right),\:{Y}_{i}^{\left(m\right)}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{i}^{\left(m\right)}$$\end{document} denotes the raw record (including grades, exam scores, and related fields) for student i in dataset m. The key design principle is that, across all three datasets, the semantics of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}^{\left(m\right)}=1$$\end{document} are aligned to “performance below an acceptable threshold” or “high risk of poor outcome,” so that the FL framework always tackles a conceptually consistent binary at-risk prediction task. For the primary dataset (Student Performance and Learning Behavior, hereafter m=1), the raw performance is captured by variables such as a categorical final grade \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{i}^{\left(1\right)}\in\:\{\mathrm{0,1},\mathrm{2,3}\}$$\end{document} (with 0 representing the lowest performance band and 3 the highest) and a continuous exam score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}^{\left(1\right)}$$\end{document}. In this work, we define a discrete performance score5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{i}^{\left(1\right)}={G}_{i}^{\left(1\right)}$$\end{document} and choose a threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{1}$$\end{document} on this ordinal scale to distinguish at-risk from non-at-risk students. A natural choice, used in our experiments, is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{1}=1$$\end{document}, which yields6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}^{\left(1\right)}=\left\{\begin{array}{ll}1,&\:\mathrm{\:if\:}{S}_{i}^{\left(1\right)}\le\:{\tau\:}_{1},\\\:0,&\:\mathrm{\:if\:}{S}_{i}^{\left(1\right)}>{\tau\:}_{1}.\end{array}\right.$$\end{document} This mapping treats students in the lowest and lower-middle grade categories (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{\mathrm{0,1}\right\}$$\end{document}) as at risk and those in higher categories (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{\mathrm{2,3}\right\}$$\end{document}) as not at risk. The threshold can also be calibrated to match a desired at-risk prevalence (for example by aligning with institutional definitions or selecting a particular quantile of the grade distribution), but the above formulation provides a clear and reproducible rule. For the Students’ Academic Performance (xAPI-Edu-Data, m=2) dataset, the raw performance is encoded directly as a categorical variable \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{i}^{\left(2\right)}\in\:\{\mathrm{L},\mathrm{M},\mathrm{H}\}$$\end{document}, denoting low, medium, and high academic performance, respectively. In this case, the mapping to the at-risk label is straightforward and does not require an additional threshold on a continuous scale:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}^{\left(2\right)}=\left\{\begin{array}{ll}1,&\:\mathrm{\:if\:}{C}_{i}^{\left(2\right)}=\mathrm{L}\\\:0,&\:\mathrm{\:if\:}{C}_{i}^{\left(2\right)}\in\:\{\mathrm{M},\mathrm{H}\}\end{array}\right.$$\end{document} This definition is consistent with the educational intuition that only the lowest performance category should be flagged as at risk, while medium and high-performance levels are treated as acceptable or safe. It also ensures that the label construction remains purely categorical, preserving the original semantics of the dataset. For the Students Performance in Exams dataset ( m=3), student achievement is represented by three exam scores: mathematics \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{M}_{i}^{\left(3\right)}$$\end{document}, reading \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}_{i}^{\left(3\right)}$$\end{document}, and writing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{i}^{\left(3\right)}$$\end{document}. We first aggregate these into a composite exam performance score, for example as a simple arithmetic mean:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{i}^{\left(3\right)}=\frac{1}{3}\left({M}_{i}^{\left(3\right)}+{R}_{i}^{\left(3\right)}+{W}_{i}^{\left(3\right)}\right)$$\end{document} We then define an at-risk threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{3}$$\end{document} on this continuous score. In our setting, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{3}$$\end{document} can be chosen either as a fixed educational cutoff (e.g., 60 out of 100) or as a data-driven percentile (e.g., the q-th quantile of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{S}_{i}^{\left(3\right)}\right\}$$\end{document}). The binary label is then:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}^{\left(3\right)}=\left\{\begin{array}{ll}1,&\:\mathrm{\:if\:}{S}_{i}^{\left(3\right)}<{\tau\:}_{3},\\\:0,&\:\mathrm{\:if\:}{S}_{i}^{\left(3\right)}\ge\:{\tau\:}_{3}.\end{array}\right.$$\end{document} This approach yields a clear notion of “underperforming across core subjects” and allows the fraction of at-risk students to be tuned via \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\tau\:}_{3}$$\end{document} in line with realistic risk prevalence. Given these label definitions, the prediction target of the framework in all datasets is the conditional probability of being at risk given the feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}^{\left(m\right)}$$\end{document}. In this study, this probability is estimated using the global logistic-regression classifier defined in Table 2. For dataset m, we write:10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}^{\left(m\right)}={f}_{\theta\:}^{\left(m\right)}\left({x}_{i}^{\left(m\right)}\right)=\sigma\:\left({{w}^{\left(m\right)}}^{\mathrm{T\:}}{x}_{i}^{\left(m\right)}+{b}^{\left(m\right)}\right)\approx\:P\left({Y}_{i}^{\left(m\right)}=1\mid\:{X}_{i}^{\left(m\right)}={x}_{i}^{\left(m\right)}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(m\right)}=({w}^{\left(m\right)},{b}^{\left(m\right)})$$\end{document} denotes the logistic-regression parameters for dataset m, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}^{\left(m\right)}$$\end{document} is the feature-weight vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}^{\left(m\right)}$$\end{document}is the intercept term, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:(\cdot\:)$$\end{document} is the logistic sigmoid function. A hard prediction is obtained using a decision threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:\in\:\left(\mathrm{0,1}\right)$$\end{document}:11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{Y}}_{i}^{\left(m\right)}=\left\{\begin{array}{cc}1,&\:\mathrm{if\:}{p}_{i}^{\left(m\right)}\ge\:\tau\:,\\\:0,&\:\mathrm{if\:}{p}_{i}^{\left(m\right)}<\tau\:.\end{array}\right.$$\end{document} For evaluation, we also use the continuous probability scores \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}^{\left(m\right)}$$\end{document} to compute ranking-based metrics such as the area under the curve (AUC) or area under the receiver operating characteristic curve. By explicitly specifying \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}^{\left(m\right)}(\cdot\:)$$\end{document} for each dataset and unifying the prediction target as the probability of being at risk, this subsection establishes a coherent and mathematically precise supervised-learning formulation for the subsequent federated, privacy-preserving, and verifiable training procedures. Data preprocessing and school simulation We describe the preprocessing pipeline applied to the primary dataset and the procedure used to simulate multiple schools from a single centralized collection of student records43. Let each raw student record be denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{i}\in\:{\mathbb{R}}^{{d}_{\mathrm{naw\:}}}$$\end{document}, comprising all available demographic, behavioral, and performance-related attributes. These raw attributes are transformed into a standardized feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}\in\:{\mathbb{R}}^{d}$$\end{document} through a preprocessing function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\phi\:(\cdot\:)$$\end{document}, i.e.,12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}=\phi\:\left({r}_{i}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\phi\:(\cdot\:)$$\end{document} includes (1) handling of missing values, (2) encoding of categorical features (e.g., one-hot encoding of gender, learning style, internet access, etc.), and (3) scaling of numerical variables (e.g., study hours, attendance, exam score, stress level) using standardization or min-max normalization. After this step, the primary dataset can be represented as13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{D}={\left\{\left({x}_{i},{y}_{i}\right)\right\}}_{i=1}^{N}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} is the at-risk label defined as in Sect. 3.2.1. Table 1Simulated school partitions on the primary dataset (student performance and learning behavior).School IDNumber of samples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\boldsymbol{n}}_{\boldsymbol{k}}\right)$$\end{document}At-risk rate (%)Mean exam score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\bar{{E}}}_{{k}}$$\end{document}Mean final grade \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\bar{{G}}}_{{k}}$$\end{document}Dominant behavior profileNotes on partitioning criteriaS11,42022.571.32.1High attendance, moderate study hoursCluster centered on high attendance and regular resource usageS21,36528.767.81.9Lower study hours, frequent use of online resourcesCluster with intensive e-learning but below-average study durationS31,51031.265.41.8High stress level, low assignment completionStudents reporting high stress and weak homework completionS41,33019.874.12.3Balanced engagement and attendanceMixed behavior, relatively strong overall performanceS51,40526.969.22.0Medium attendance, high use of learning technologiesCluster with frequent EduTech usage and moderate exam scoresS61,36033.563.91.7Low study hours, irregular participation in discussionsUnder-engaged students with sporadic participationS71,41524.172.02.2High study hours, regular assignment completionStrong academic habits and consistent homework completionS81,36529.468.31.9Mixed attendance, medium motivationHeterogeneous group with moderate motivation levelsS91,41027.868.92.0Frequent use of online courses and digital resourcesStudents relying heavily on online courses and materialsS101,42325.670.52.1Balanced behavior, average stress levelResidual cluster with average behavioral and performance patternsTotal14,003----Entire primary dataset (sum over all simulated schools) Table 1Simulated school partitions on the primary dataset (student performance and learning behavior).School IDNumber of samples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left({\boldsymbol{n}}_{\boldsymbol{k}}\right)$$\end{document}At-risk rate (%)Mean exam score \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\bar{{E}}}_{{k}}$$\end{document}Mean final grade \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\bar{{G}}}_{{k}}$$\end{document}Dominant behavior profileNotes on partitioning criteriaS11,42022.571.32.1High attendance, moderate study hoursCluster centered on high attendance and regular resource usageS21,36528.767.81.9Lower study hours, frequent use of online resourcesCluster with intensive e-learning but below-average study durationS31,51031.265.41.8High stress level, low assignment completionStudents reporting high stress and weak homework completionS41,33019.874.12.3Balanced engagement and attendanceMixed behavior, relatively strong overall performanceS51,40526.969.22.0Medium attendance, high use of learning technologiesCluster with frequent EduTech usage and moderate exam scoresS61,36033.563.91.7Low study hours, irregular participation in discussionsUnder-engaged students with sporadic participationS71,41524.172.02.2High study hours, regular assignment completionStrong academic habits and consistent homework completionS81,36529.468.31.9Mixed attendance, medium motivationHeterogeneous group with moderate motivation levelsS91,41027.868.92.0Frequent use of online courses and digital resourcesStudents relying heavily on online courses and materialsS101,42325.670.52.1Balanced behavior, average stress levelResidual cluster with average behavioral and performance patternsTotal14,003----Entire primary dataset (sum over all simulated schools) Simulated school partitions on the primary dataset (student performance and learning behavior). To emulate a multi-school federated setting, the centralized dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{D}$$\end{document} is partitioned into K=10 disjoint subsets, each interpreted as a distinct school participating as a federated client. Formally, we define a school assignment function14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:s:\{1,\dots\:,N\}\to\:\{1,\dots\:,K\}$$\end{document} and construct the local dataset of school k as:15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}=\left\{\left({x}_{i},{y}_{i}\right)\in\:\mathcal{D}\mid\:s\left(i\right)=k\right\},\:k=1,\dots\:,K$$\end{document} The subsets \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}$$\end{document} are by construction disjoint and cover the full dataset, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{D}=\bigcup\:_{k=1}^{K}\:{\mathcal{D}}_{k}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}\cap\:{\mathcal{D}}_{j}={\varnothing}$$\end{document} for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\ne\:j$$\end{document}. In practice, the assignment function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:s(\cdot\:)$$\end{document} is derived by grouping students with similar behavioral and performance profiles (e.g., attendance, study hours, assignment completion, use of digital resources, stress level), so that each resulting group corresponds to a “school” with internally coherent characteristics and externally distinct distributions. This induces a heterogeneous (non-IID) federated scenario in which schools differ in size, average performance, and at-risk prevalence. For the primary dataset, the school assignment function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:s(\cdot\:)$$\end{document} was implemented using K-means clustering with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{1}=10$$\end{document}. Clustering was performed after preprocessing and standardizing the behavioral and contextual variables, including Attendance, StudyHours, AssignmentCompletion, EduTech usage, Discussions, StressLevel, Motivation, LearningStyle, and Internet access. Performance variables directly related to the label definition, such as FinalGrade and ExamScore, were not used for assigning students to schools; they were used only for descriptive reporting of the resulting partitions. K-means was initialized with k-means + + and run with a fixed random seed of 1 to ensure reproducibility. Each resulting cluster was treated as one simulated school. For each simulated school k, we report in Table 1 the number of samples \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{k}=\left|{\mathcal{D}}_{k}\right|$$\end{document}, the at-risk rate, and summary performance statistics. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{Y}_{i}$$\end{document} denote the at-risk label and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{i}$$\end{document} denote the exam score and final grade for student i, respectively. The at-risk rate in school k is computed as16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\rho\:}_{k}=\frac{1}{{n}_{k}}\sum\:_{\left({x}_{i},{y}_{i}\right)\in\:{\mathcal{D}}_{k}}\:\mathbb{I}\left({Y}_{i}=1\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathbb{I}(\cdot\:)$$\end{document} is the indicator function. The mean exam score and mean final grade reported in Table 1 are given by:17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\bar{E}}_{k}=\frac{1}{{n}_{k}}\sum\:_{\left({x}_{i},{y}_{i}\right)\in\:{\mathcal{D}}_{k}}\:{E}_{i},\:{\bar{G}}_{k}=\frac{1}{{n}_{k}}\sum\:_{\left({x}_{i},{y}_{i}\right)\in\:{\mathcal{D}}_{k}}\:{G}_{i}$$\end{document} As shown in Table 1, the simulated schools have comparable but not identical sizes (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{k}$$\end{document} ranging from 1,330 to 1,510 students), and display substantial variation in at-risk rates (from 19.8% in S4 to 33.5% in S6) as well as in mean exam scores and final grades. The behavioral descriptions further highlight qualitative differences between schools, such as clusters characterized by high attendance and strong study habits (S1, S7) versus those with low study hours or irregular participation (S2, S6). This heterogeneity in both outcome prevalence and behavioral patterns is desirable for the subsequent FL experiments, as it induces a non-IID client environment that more closely mirrors real-world multi-school deployments. Extending the federated scenario to the additional datasets We describe how the FL scenario defined on the primary dataset is extended to the two additional datasets while keeping a single, consistent prediction model across all experimental settings. For each dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\in\:\{\mathrm{1,2},3\}$$\end{document}, let the preprocessed student records be represented as:18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}^{\left(m\right)}={\left\{\left({x}_{i}^{\left(m\right)},{y}_{i}^{\left(m\right)}\right)\right\}}_{i=1}^{{N}_{m}}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}^{\left(m\right)}\in\:{\mathbb{R}}^{{d}^{\left(m\right)}}$$\end{document} denotes the preprocessed feature vector obtained from dataset-specific cleaning, encoding, and scaling, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}^{\left(m\right)}\in\:\left\{\mathrm{0,1}\right\}$$\end{document} denotes the binary at-risk label defined in Sect. 3.2.1. Each dataset is partitioned into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{m}$$\end{document}disjoint subsets,19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left\{{\mathcal{D}}_{k}^{\left(m\right)}\right\}}_{k=1}^{{K}_{m}}$$\end{document} where each subset represents a simulated school or federated client. The local dataset of client k in dataset m is defined as:20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}^{\left(m\right)}=\left\{\left({x}_{i}^{\left(m\right)},{y}_{i}^{\left(m\right)}\right)\in\:{\mathcal{D}}^{\left(m\right)}\mid\:{s}^{\left(m\right)}\left(i\right)=k\right\},k=1,\dots\:,{K}_{m}$$\end{document} The resulting client datasets satisfy:21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}^{\left(m\right)}=\bigcup\:_{k=1}^{{K}_{m}}{\mathcal{D}}_{k}^{\left(m\right)},{\mathcal{D}}_{k}^{\left(m\right)}\cap\:{\mathcal{D}}_{j}^{\left(m\right)}={\varnothing}\mathrm{for\:}k\ne\:j$$\end{document} The assignment function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{s}^{\left(m\right)}(\cdot\:)$$\end{document} is defined separately for each dataset according to the available institutional or cohort-level structure. For the primary dataset, students are grouped into simulated schools using behavioral and performance-related profiles, as described in Sect. 3.2.2. For the xAPI-Edu-Data dataset, school partitions are derived from available grade and section identifiers, with small groups merged when necessary to avoid extremely small clients. For the Students Performance in Exams dataset, where no natural school identifier is available, stratified random partitioning is used to preserve the at-risk ratio across simulated schools. These partitioning strategies are designed to induce heterogeneous, non-IID client distributions while keeping the same FL framework across datasets. Across all datasets, schools, and training regimes, we use a single global logistic-regression classifier for binary at-risk prediction. The same model form is used in the centralized baseline, standard FL, and the proposed differentially private FL with PBT. This design choice avoids confounding the comparison across training regimes with changes in model architecture or capacity. For a given dataset m, the model maps a student feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}^{\left(m\right)}$$\end{document} to an estimated probability of being at risk as follows:22\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}^{\left(m\right)}={f}_{\theta\:}^{\left(m\right)}\left({x}_{i}^{\left(m\right)}\right)=\sigma\:\left({{w}^{\left(m\right)}}^{\mathrm{T\:}}{x}_{i}^{\left(m\right)}+{b}^{\left(m\right)}\right)\approx\:P\left({Y}_{i}^{\left(m\right)}=1\mid\:{X}_{i}^{\left(m\right)}={x}_{i}^{\left(m\right)}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(m\right)}=({w}^{\left(m\right)},{b}^{\left(m\right)})$$\end{document} denotes the model parameters for dataset m, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}^{\left(m\right)}\in\:{\mathbb{R}}^{{d}^{\left(m\right)}}$$\end{document} is the feature-weight vector, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}^{\left(m\right)}$$\end{document} is the intercept term, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:(\cdot\:)$$\end{document} is the logistic sigmoid function,23\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\left(z\right)=\frac{1}{1+\mathrm{e}\mathrm{x}\mathrm{p}(-z)}$$\end{document} The parameters are optimized by minimizing the empirical binary cross-entropy loss. For dataset m, the centralized objective is:24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}^{\left(m\right)}\left(\theta\:\right)=-\frac{1}{{N}_{m}}\sum\:_{i=1}^{{N}_{m}}\left[{y}_{i}^{\left(m\right)}log{p}_{i}^{\left(m\right)}+\left(1 -{y}_{i}^{\left(m\right)}\right)log\left(1- {p}_{i}^{\left(m\right)}\right)\right]$$\end{document} In the federated setting, each school kminimizes the same binary cross-entropy objective on its local dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}^{\left(m\right)}$$\end{document}. The server then aggregates the locally updated logistic-regression parameters using the FedAvg procedure described in Sect. 3.3. In the differentially private setting, local gradients are clipped and perturbed with Gaussian noise before aggregation, while the PBT protocol records verifiable commitments to client participation and model state at each communication round. Table 2Specification of the global logistic-regression model used in the federated framework.ComponentSpecificationDescriptionInput \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}^{\left(m\right)}\in\:{\mathbb{R}}^{{d}^{\left(m\right)}}$$\end{document} Preprocessed numerical and one-hot encoded student features for dataset mModel typeLogistic regressionA linear probabilistic classifier shared across centralized, FL, and FL + DP+PBT settingsParameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(m\right)}=({w}^{\left(m\right)},{b}^{\left(m\right)})$$\end{document} Feature-weight vector and intercept termPrediction function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}^{\left(m\right)}=\sigma\:({{w}^{\left(m\right)}}^{\mathrm{T\:}}{x}_{i}^{\left(m\right)}+{b}^{\left(m\right)})$$\end{document} Estimated probability that student iis at riskOutput \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}^{\left(m\right)}\in\:\left(\mathrm{0,1}\right)$$\end{document} Continuous at-risk probability used for classification and AUC computationDecision rule \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{i}^{\left(m\right)}=\mathbb{I}({p}_{i}^{\left(m\right)}\ge\:\tau\:)$$\end{document} Binary at-risk prediction using threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document}, typically \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:=0.5$$\end{document}Loss functionBinary cross-entropyOptimized in centralized and local federated trainingCentralized trainingMini-batch optimization on pooled dataBaseline setting with all data available on a single serverFederated trainingFedAvg over local model parametersSchools train locally and share only model updatesDP mechanismGradient clipping + Gaussian noiseClient-side differentially private optimizationPBT mechanismRound-wise client/model commitmentVerifiable logging of client participation before local trainingInterpretabilityCoefficient-based feature importanceFeature importance obtained from learned logistic-regression weights Table 2Specification of the global logistic-regression model used in the federated framework.ComponentSpecificationDescriptionInput \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{x}_{i}^{\left(m\right)}\in\:{\mathbb{R}}^{{d}^{\left(m\right)}}$$\end{document} Preprocessed numerical and one-hot encoded student features for dataset mModel typeLogistic regressionA linear probabilistic classifier shared across centralized, FL, and FL + DP+PBT settingsParameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(m\right)}=({w}^{\left(m\right)},{b}^{\left(m\right)})$$\end{document} Feature-weight vector and intercept termPrediction function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}^{\left(m\right)}=\sigma\:({{w}^{\left(m\right)}}^{\mathrm{T\:}}{x}_{i}^{\left(m\right)}+{b}^{\left(m\right)})$$\end{document} Estimated probability that student iis at riskOutput \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}^{\left(m\right)}\in\:\left(\mathrm{0,1}\right)$$\end{document} Continuous at-risk probability used for classification and AUC computationDecision rule \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{i}^{\left(m\right)}=\mathbb{I}({p}_{i}^{\left(m\right)}\ge\:\tau\:)$$\end{document} Binary at-risk prediction using threshold \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:$$\end{document}, typically \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\tau\:=0.5$$\end{document}Loss functionBinary cross-entropyOptimized in centralized and local federated trainingCentralized trainingMini-batch optimization on pooled dataBaseline setting with all data available on a single serverFederated trainingFedAvg over local model parametersSchools train locally and share only model updatesDP mechanismGradient clipping + Gaussian noiseClient-side differentially private optimizationPBT mechanismRound-wise client/model commitmentVerifiable logging of client participation before local trainingInterpretabilityCoefficient-based feature importanceFeature importance obtained from learned logistic-regression weights Specification of the global logistic-regression model used in the federated framework. The use of logistic regression is intentional. First, it provides a transparent and auditable global model whose coefficients can be directly inspected as feature-importance indicators. Second, it reduces computational and communication overhead in the federated setting, which is important for practical deployment across schools with limited technical infrastructure. Third, by using the same simple model across centralized, federated, and privacy-preserving federated settings, the observed performance differences can be attributed to the training paradigm, privacy mechanism, and client heterogeneity rather than to differences in model complexity. The specification of the global logistic-regression model used throughout the experiments is summarized in Table 2. As shown in Table 2, the model contains no hidden layers or dataset-specific architectural components. This ensures that the comparison between centralized learning, standard FL, and the proposed FL + DP+PBT setting is not affected by changes in model capacity. Moreover, because the model is linear in the input features, the learned coefficients provide a direct and reproducible basis for the feature-importance analysis reported in the Results section. Federated learning algorithm with DP This section presents the training algorithm used in the proposed framework, combining standard federated optimization with a client-side DP mechanism for the logistic-regression model defined in Table 2. Let the primary and additional datasets, after preprocessing and school-level partitioning, be denoted generically by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{\mathcal{D}}_{k}{\}}_{k=1}^{K}\right.$$\end{document}, where:25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}={\left\{\left({x}_{i}^{\left(k\right)},{y}_{i}^{\left(k\right)}\right)\right\}}_{i=1}^{{n}_{k}}$$\end{document} is the local dataset of school k. The global model is a logistic-regression classifier with parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:=(w,b)$$\end{document}, where wis the feature-weight vector and b is the intercept term. For an input feature vector x, the model outputs the probability of being at risk as:26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{\theta\:}\left(x\right)=\sigma\:({w}^{\mathrm{T\:}}x+b)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:(\cdot\:)$$\end{document} is the logistic sigmoid function. The local objective at school kis the empirical binary cross-entropy loss:27\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{F}_{k}\left(\theta\:\right)=\frac{1}{{n}_{k}}\sum\:_{i=1}^{{n}_{k}}\mathcal{l}\left({f}_{\theta\:}\left({x}_{i}^{\left(k\right)}\right),{y}_{i}^{\left(k\right)}\right)$$\end{document} where:28\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{l}(p,y)=-\left[y\mathrm{l}\mathrm{o}\mathrm{g}p+(1-y)\mathrm{l}\mathrm{o}\mathrm{g}(1-p)\right]$$\end{document} The FL problem aims to learn a single global logistic-regression model by minimizing the weighted sum of local empirical risks across schools:29\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\underset{\theta\:}{\mathrm{m}\mathrm{i}\mathrm{n}}F\left(\theta\:\right)\triangleq\:\underset{\theta\:}{\mathrm{m}\mathrm{i}\mathrm{n}}\sum\:_{k=1}^{K}\frac{{n}_{k}}{N}{F}_{k}\left(\theta\:\right),N=\sum\:_{k=1}^{K}{n}_{k}$$\end{document} This objective is optimized under the constraint that raw student records remain stored locally at the corresponding schools and are never transferred to the central server. Training proceeds in synchronous communication rounds indexed by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:t=\mathrm{0,1},\dots\:,T-1$$\end{document}. At the beginning of round t, the server maintains a global parameter vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(t\right)}$$\end{document} and selects a subset of schools:30\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{S}}^{\left(t\right)}\subseteq\:\{1,\dots\:,K\}$$\end{document} to participate in the round, for example by uniform sampling with participation rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left.{C}_{\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{t}}\in\:(\mathrm{0,1}\right]$$\end{document}. The server broadcasts \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(t\right)}$$\end{document} to all selected schools. Each selected school \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\in\:{\mathcal{S}}^{\left(t\right)}$$\end{document} initializes a local copy:31\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k}^{\left(t,0\right)}={\theta\:}^{\left(t\right)}$$\end{document} and performs \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:E\ge\:1$$\end{document} local epochs of mini-batch gradient-based optimization on its local objective. For a mini-batch \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{B}_{k}^{\left(t,e\right)}\subset\:{\mathcal{D}}_{k}$$\end{document} at local step \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(t,e\right)$$\end{document}, the non-private local update is:32\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k}^{\left(t,e+1\right)}={\theta\:}_{k}^{\left(t,e\right)}-\eta\:\frac{1}{\mid\:{B}_{k}^{\left(t,e\right)}\mid\:}\sum\:_{(x,y)\in\:{B}_{k}^{\left(t,e\right)}}{\nabla\:}_{\theta\:}\mathcal{l}\left({f}_{{\theta\:}_{k}^{\left(t,e\right)}}\left(x\right),y\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:>0$$\end{document} is the local learning rate and eindexes local update steps. After completing the prescribed local training, school kobtains a locally updated parameter vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k}^{\left(t+1\right)}$$\end{document}. In the non-private FedAvg scheme, each selected school sends \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k}^{\left(t+1\right)}$$\end{document} to the server, and the server aggregates the received model parameters using sample-size weighting:33\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(t+1\right)}=\sum\:_{k\in\:{\mathcal{S}}^{\left(t\right)}}{\alpha\:}_{k}^{\left(t\right)}{\theta\:}_{k}^{\left(t+1\right)}$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{k}^{\left(t\right)}$$\end{document} is the aggregation weight of client k at round t, computed as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{k}^{\left(t\right)}={n}_{k}/{\sum\:}_{j\in\:{\mathcal{S}}^{\left(t\right)}}{n}_{j}$$\end{document}. This aggregation rule produces a new global logistic-regression model after each communication round. The procedure is repeated for T rounds, producing a sequence of global models:34\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\left\{{\theta\:}^{\left(t\right)}\right\}}_{t=0}^{T}$$\end{document} The centralized baseline used for comparison corresponds to optimizing the same binary cross-entropy objective directly on the pooled training dataset:35\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{D}=\bigcup\:_{k=1}^{K}{\mathcal{D}}_{k}$$\end{document} To preserve the privacy of individual students in the federated setting, we integrate a client-side DP mechanism based on the DP-SGD paradigm. We adopt the standard notion of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\epsilon\:,\delta\:\right)$$\end{document}-DP: a randomized mechanism \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathcal{M}$$\end{document} acting on datasets satisfies \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\epsilon\:,\delta\:\right)$$\end{document}-DP if, for any pair of neighbouring datasets Dand \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{D}^{{\prime\:}}$$\end{document}differing in one student record and for any measurable set of outputs S,36\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Pr\left[\mathcal{M}\right(D)\in\:S]\le\:{e}^{\epsilon\:}Pr\left[\mathcal{M}\right({D}^{{\prime\:}})\in\:S]+\delta\:$$\end{document} In the proposed framework, each school implements a local randomized mechanism that maps its local dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{k}$$\end{document} to a noisy model update before transmission to the server. For each mini-batch \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{B}_{k}^{\left(t,e\right)}$$\end{document}, school kfirst computes per-example gradients with respect to the logistic-regression parameters:37\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{i}^{\left(t,e\right)}={\nabla\:}_{\theta\:}\mathcal{l}\left({f}_{{\theta\:}_{k}^{\left(t,e\right)}}\left({x}_{i}^{\left(k\right)}\right),{y}_{i}^{\left(k\right)}\right),\left({x}_{i}^{\left(k\right)},{y}_{i}^{\left(k\right)}\right)\in\:{B}_{k}^{\left(t,e\right)}$$\end{document} Each per-example gradient is clipped to a fixed \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{l}}_{2}$$\end{document}-norm bound \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{p}}>0$$\end{document}:38\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\prime }{g}}_{i}^{\left(t,e\right)}={g}_{i}^{\left(t,e\right)}\cdot\:\mathrm{m}\mathrm{i}\mathrm{n}\left(1,\frac{{C}_{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{p}}}{{\parallel\:{g}_{i}^{\left(t,e\right)}\parallel\:}_{2}}\right)$$\end{document} The clipped gradients are then averaged over the mini-batch:39\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\prime }{g}}_{k}^{\left(t,e\right)}=\frac{1}{\mid\:{B}_{k}^{\left(t,e\right)}\mid\:}\sum\:_{\left({x}_{i}^{\left(k\right)},{y}_{i}^{\left(k\right)}\right)\in\:{B}_{k}^{\left(t,e\right)}}{\stackrel{\prime }{g}}_{i}^{\left(t,e\right)}$$\end{document} Gaussian noise calibrated to the clipping bound is added to the averaged clipped gradient:40\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\stackrel{\sim}{g}}_{k}^{\left(t,e\right)}={\stackrel{\prime }{g}}_{k}^{\left(t,e\right)}+\mathcal{N}\left(0,{\sigma\:}^{2}{C}_{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{p}}^{2}I\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:>0$$\end{document} is the noise multiplier and Iis the identity matrix with the same dimension as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:$$\end{document}. The private local update is then performed as:41\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k}^{\left(t,e+1\right)}={\theta\:}_{k}^{\left(t,e\right)}-\eta\:{\stackrel{\sim}{g}}_{k}^{\left(t,e\right)}$$\end{document} After all local steps in round tare completed, the final local parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k}^{\left(t+1\right)}$$\end{document} incorporate the noise added during local training. These noisy local parameters are transmitted to the server and aggregated using the same FedAvg rule:42\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(t+1\right)}=\sum\:_{k\in\:{\mathcal{S}}^{\left(t\right)}}{\alpha\:}_{k}^{\left(t\right)}{\theta\:}_{k}^{\left(t+1\right)}$$\end{document} where:43\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\alpha\:}_{k}^{\left(t\right)}=\frac{{n}_{k}}{\sum\:_{j\in\:{\mathcal{S}}^{\left(t\right)}}{n}_{j}}$$\end{document} From the server’s perspective, the aggregation step remains unchanged; however, the local updates have already been randomized at the client side before leaving each school. Under standard DP-SGD accounting assumptions, the sequence of clipped and noisy local updates yields an overall \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\epsilon\:,\delta\:\right)$$\end{document}-DP guarantee for each student’s contribution to the local training data. The effective privacy budget \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\epsilon\:$$\end{document}is controlled by the noise multiplier \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}, the clipping bound \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{p}}$$\end{document}, the mini-batch size, the number of local optimization steps, and the number of communication rounds. In the experiments, multiple noise multipliers are considered to examine the privacy–utility trade-off in terms of both predictive performance and the resulting privacy budget. This DP-FL procedure is used consistently across the three educational datasets and the corresponding simulated school partitions. Because the underlying classifier remains the same logistic-regression model in all centralized, federated, and differentially private settings, the observed differences in performance can be attributed to the effects of decentralization, non-IID client distributions, and privacy-preserving noise injection rather than to changes in model architecture. The procedure is also compatible with the PBT protocol described in Sect. 3.4. The cryptographic commitments used in that protocol operate only on model states, client identifiers, and round metadata, while the DP mechanism ensures that the model updates transmitted from schools are already randomized before aggregation. As a result, the framework combines decentralized optimization, client-side DP, and verifiable participation logging within a single federated at-risk prediction pipeline. Proof-before-train protocol The PBT protocol is designed to provide a verifiable record of client participation in each federated communication round before any local training is performed. In the proposed framework, this mechanism does not modify the learning objective or the DP mechanism. Instead, it operates as an auditability layer on top of federated training by binding each participating school to a specific global model state and communication round before local optimization begins. At the beginning of round t, the server broadcasts the current global logistic-regression parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(t\right)}$$\end{document} to the selected set of schools \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{S}}^{\left(t\right)}$$\end{document}. Before performing local training, each participating school \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:k\in\:{\mathcal{S}}^{\left(t\right)}$$\end{document} computes a cryptographic commitment:44\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Gamma\:}}_{k}^{\left(t\right)}=H\left({\theta\:}^{\left(t\right)}\left\| k \right. \left\| t \right. \left\|{r}_{k}\right.^{\left(t\right)}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H(\cdot\:)$$\end{document} denotes a cryptographic hash function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(t\right)}$$\end{document} is the current global model state, kis the client identifier, tis the communication-round index, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{k}^{\left(t\right)}$$\end{document} is a client-generated nonce used to prevent replay or reuse of old commitments. The school sends \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Gamma\:}}_{k}^{\left(t\right)}$$\end{document} to the server before starting local training. The server stores this commitment together with the round metadata as part of the audit log. After the commitment has been registered, school kperforms local training on its own data using either the non-private FedAvg update or the client-side DP-SGD update described in Sect. 3.3. When the local update is completed, the school sends the updated model parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{k}^{\left(t+1\right)}$$\end{document}, together with the corresponding metadata needed to verify the previously submitted commitment. The server verifies that the received update is associated with the same client, communication round, and global model state for which the pre-training commitment was registered. Algorithm 1 summarizes the PBT procedure executed at each federated communication round. The key idea is that each selected school must register a cryptographic commitment to the current global model state before performing any local training. This commitment binds the client identity, communication-round index, global logistic-regression parameters, and a fresh client-generated nonce into a single hash value. After local training, the server verifies that the submitted update corresponds to the previously registered commitment before accepting it for aggregation. In this way, PBT provides an audit trail of client participation and model-state consistency without requiring access to raw student records. The protocol operates only on model parameters and round metadata, and therefore does not interfere with the client-side DP-SGD mechanism used to protect individual student records. Algorithm 1PBT protocol for round t. PBT protocol for round t. This procedure provides three auditability properties. First, it records which schools participated in each communication round. Second, it binds each local update to the global model state distributed at the beginning of that round. Third, it reduces the risk of unregistered or post-hoc updates being accepted by the server. Importantly, the protocol does not require access to raw student records and does not weaken the DP mechanism, because commitments are computed only from model states and metadata. The PBT protocol is lightweight. Its computational cost is dominated by hashing model-state metadata once per participating client and communication round, which is negligible compared with local model optimization. Therefore, the protocol improves verifiability and accountability in the federated training process while introducing minimal computational overhead. Results In this section, we present the experimental results of the proposed framework on the three education datasets. We first compare the centralized baseline, vanilla FL, and FL with DP on the main dataset, with regards to classification performance, convergence behavior, and privacy–utility trade-off. Subsequently, the same set of experiments are repeated on the two supplementary datasets. We then evaluate the impact of the PBT protocol on test accuracy, stability over communication rounds, and computational overhead. Experimental setup We evaluate the proposed framework on three educational datasets after transforming them into a unified binary at-risk prediction task, where y=1denotes at-risk students, i.e., students with low performance or performance below a predefined threshold, and y=0 denotes not-at-risk students, i.e., students with acceptable or high performance, as defined in Sect. 3.2.1. For each dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:m\in\:\{\mathrm{1,2},3\}$$\end{document}, the preprocessed data are denoted by:45\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}^{\left(m\right)}=\left\{\right({x}_{i}^{\left(m\right)},{y}_{i}^{\left(m\right)}){\}}_{i=1}^{{N}_{m}}$$\end{document} We perform a stratified split into training, validation, and test sets with proportions 70%, 10%, and 20%, respectively, preserving the at-risk class ratio in each split. The same splits are used for all training regimes, including centralized learning, standard FL, and differentially private FL with PBT. The test set \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{D}}_{\mathrm{t}\mathrm{e}\mathrm{s}\mathrm{t}}^{\left(m\right)}$$\end{document} is held out entirely for final evaluation. On top of this global split, we simulate a multi-school federated environment by partitioning the data into disjoint school-level subsets. For the primary dataset, Student Performance and Learning Behavior \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(m=1\right)$$\end{document}, the training, validation, and test sets are each partitioned into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{1}=10$$\end{document} schools using a clustering-based assignment that groups students with similar behavioral and performance profiles, as described in Sect. 3.2.2. For the xAPI-Edu-Data dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(m=2,{N}_{2}\approx\:480\right)$$\end{document}, we define \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{2}=5$$\end{document} schools based on merged grade and section identifiers to avoid very small clients. For the Students Performance in Exams dataset \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(m=3,{N}_{3}=1000\right)$$\end{document}, we use \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{K}_{3}=5$$\end{document} schools obtained through stratified random partitioning that preserves the at-risk ratio within each school. Federated training always uses the union of all school-level training subsets, while performance is measured on the global test set aggregated across schools, in order to assess system-level generalization rather than performance for a single institution. All experiments, centralized and federated, use the same global logistic-regression model described in Table 2, with the input dimension adjusted to the feature space of each dataset. The model outputs an at-risk probability:46\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{p}={f}_{\theta\:}\left(x\right)=\sigma\:({w}^{\mathrm{T\:}}x+b)\in\:\left(\mathrm{0,1}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\theta\:=(w,b)$$\end{document} denotes the logistic-regression parameters. The model is trained with binary cross-entropy loss. Using the same logistic-regression model across all settings ensures that performance differences between centralized learning, standard FL, and FL + DP+PBT are attributable to the training paradigm, client heterogeneity, and privacy mechanism rather than to changes in model architecture or capacity. The centralized baseline is trained on the pooled training data using mini-batch optimization with learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{-3}$$\end{document}, batch size 64, weight decay \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{-5}$$\end{document}, and early stopping based on validation loss with a patience of 10 epochs. The federated setting employs FedAvg: in each communication round, a fraction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{t}}=0.6$$\end{document} of schools is sampled, each selected school trains the logistic-regression model locally for E=3 epochs using learning rate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:5\times\:{10}^{-4}$$\end{document} and batch size 32, and the server aggregates the updated model parameters with sample-size weighting. We run T=100 global communication rounds, monitor validation performance periodically, and repeat all experiments with five random seeds. The main hyperparameters, including learning rate, local epochs, batch size, weight decay, client participation rate, and number of communication rounds, are tuned on the primary dataset and then transferred to the two additional datasets with only minimal adjustment when necessary. For the differentially private experiments (FL + DP), we adopt client-side DP-SGD. Per-example gradients are clipped to an \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{l}}_{2}$$\end{document}-norm bound \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{p}}=1.0$$\end{document} and perturbed with Gaussian noise of variance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}^{2}{C}_{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{p}}^{2}$$\end{document}, with noise multipliers \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:\in\:\left\{\mathrm{0.0,0.5,1.0}\right\}$$\end{document}. The case \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=0$$\end{document} recovers the non-private FL baseline, while larger values of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}provide stronger privacy at the cost of potentially reduced predictive utility. The overall privacy budget \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left(\epsilon\:,\delta\:\right)$$\end{document} is tracked using a moments accountant with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\delta\:={10}^{-5}$$\end{document}. Finally, the PBT protocol is applied to the proposed FL + DP+PBT setting and to the corresponding ablation analysis. At the start of each communication round, each participating school submits a cryptographic commitment to the current global model state, its client identifier, and the round index:47\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Gamma\:}}_{k}^{\left(t\right)}=H\left({\theta\:}^{\left(t\right)} \left\| k \right. \left\| t \right. \left\| {r}_{k} \right.^{\left(t\right)}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{k}^{\left(t\right)}$$\end{document} is a client-generated nonce, as defined in Sect. 3.4, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H(\cdot\:)$$\end{document} denotes a cryptographic hash function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}^{\left(t\right)}$$\end{document} is the current global logistic-regression parameter vector, k is the client identifier, and t is the communication-round index. The server verifies the submitted commitment when receiving the corresponding local model update. This mechanism provides a verifiable record of client participation and model state before local training, without requiring access to raw student data. This experimental setup jointly evaluates the three core components of the proposed framework: decentralized learning across heterogeneous schools, client-side DP, and PBT-based auditability. By using the same logistic-regression model across all training regimes, the comparison isolates the effects of federated optimization, privacy-preserving noise injection, and verifiable client participation on at-risk student prediction. Performance on the primary dataset In this subsection, we report and analyze the performance of the proposed framework on the primary merged dataset of 14,003 students, which serves as the main development and evaluation cohort. The prediction task is a binary at-risk classification problem, where students with FinalGrade ≤ 1 are labeled as at-risk and those with FinalGrade ∈ {2, 3} as not at-risk. Table 3 summarizes three key configurations: a centralized logistic regression baseline (M1), a federated counterpart without DP (M2), and the full proposed FL-AtRisk-DP-PBT framework (M3), which combines FL, client-side DP-SGD, and the PBT protocol. All results are reported on a held-out test set of 2,800 students obtained from a stratified 70/10/20 train/validation/test split. Table 3Configuration and test performance on the primary dataset (merged 14,003-record cohort, binary at-risk prediction, 70/10/20 split).AspectM1 – Centralized logistic regressionM2 – Federated LR (FedAvg, no DP)M3 – Proposed FL-AtRisk-DP-PBT (FedAvg + DP + PBT)Training modeCentralized (single server)Cross-school federated learningCross-school federated learningData locationAll 14,003 records on one serverData kept locally at simulated schoolsData kept locally at simulated schoolsTarget taskBinary at-risk prediction (FinalGrade ≤ 1 vs. ≥ 2)SameSameModel architectureLogistic regression (input features + intercept)Same global model as M1Same global model as M1Number of schools (clients) K–10 simulated schools10 simulated schoolsSchool partitioning–K-means on behavior/performance featuresK-means on behavior/performance featuresClient fraction per round C–0.6 (≈ 60% of schools participate each round)0.6 (≈ 60% of schools participate each round)Local epochs E–3 epochs per selected client per round3 epochs per selected client per roundLocal batch size B64 (mini-batch SGD)3232Global communication rounds T–100100OptimizerMini-batch optimization of logistic lossLocal mini-batch optimization of logistic loss with FedAvg aggregationClient-side DP optimization of logistic loss with gradient clipping and Gaussian noiseLearning rate (central/local)1 × 10−3 (centralized)5 × 10−4 (local clients)5 × 10−4 (local clients)Differential privacy (DP)NoneNoneClient-side DP-SGD with Gaussian noiseDP noise multiplier σ–0.0 (non-private)0.5 (moderate noise)Gradient clipping norm C––1.0 (per-example ℓ₂ clipping)Proof-before-train protocol (PBT)NoNoYes (cryptographic commitment per round and client)Test set size2,800 students2,800 students2,800 studentsTest accuracy0.9914 (99.14%)0.9861 (98.61%)0.9800 (98.00%)Test precision (at-risk = 1)0.99790.99220.9879Test recall (at-risk = 1)0.98530.98040.9727Test F1-score (at-risk = 1)0.99150.98630.9802Test AUC0.99980.99930.9992Δ accuracy vs. centralized (percentage points)0.00−0.53−1.14Δ F1-score vs. centralized (percentage points)0.00−0.52−1.13InterpretationUpper-bound with full centralization and no privacySmall loss due to decentralization and non-IID schoolsAdditional small loss due to DP, with formal privacy + PBT Table 3Configuration and test performance on the primary dataset (merged 14,003-record cohort, binary at-risk prediction, 70/10/20 split).AspectM1 – Centralized logistic regressionM2 – Federated LR (FedAvg, no DP)M3 – Proposed FL-AtRisk-DP-PBT (FedAvg + DP + PBT)Training modeCentralized (single server)Cross-school federated learningCross-school federated learningData locationAll 14,003 records on one serverData kept locally at simulated schoolsData kept locally at simulated schoolsTarget taskBinary at-risk prediction (FinalGrade ≤ 1 vs. ≥ 2)SameSameModel architectureLogistic regression (input features + intercept)Same global model as M1Same global model as M1Number of schools (clients) K–10 simulated schools10 simulated schoolsSchool partitioning–K-means on behavior/performance featuresK-means on behavior/performance featuresClient fraction per round C–0.6 (≈ 60% of schools participate each round)0.6 (≈ 60% of schools participate each round)Local epochs E–3 epochs per selected client per round3 epochs per selected client per roundLocal batch size B64 (mini-batch SGD)3232Global communication rounds T–100100OptimizerMini-batch optimization of logistic lossLocal mini-batch optimization of logistic loss with FedAvg aggregationClient-side DP optimization of logistic loss with gradient clipping and Gaussian noiseLearning rate (central/local)1 × 10−3 (centralized)5 × 10−4 (local clients)5 × 10−4 (local clients)Differential privacy (DP)NoneNoneClient-side DP-SGD with Gaussian noiseDP noise multiplier σ–0.0 (non-private)0.5 (moderate noise)Gradient clipping norm C––1.0 (per-example ℓ₂ clipping)Proof-before-train protocol (PBT)NoNoYes (cryptographic commitment per round and client)Test set size2,800 students2,800 students2,800 studentsTest accuracy0.9914 (99.14%)0.9861 (98.61%)0.9800 (98.00%)Test precision (at-risk = 1)0.99790.99220.9879Test recall (at-risk = 1)0.98530.98040.9727Test F1-score (at-risk = 1)0.99150.98630.9802Test AUC0.99980.99930.9992Δ accuracy vs. centralized (percentage points)0.00−0.53−1.14Δ F1-score vs. centralized (percentage points)0.00−0.52−1.13InterpretationUpper-bound with full centralization and no privacySmall loss due to decentralization and non-IID schoolsAdditional small loss due to DP, with formal privacy + PBT Configuration and test performance on the primary dataset (merged 14,003-record cohort, binary at-risk prediction, 70/10/20 split). From Table 3, model M1 establishes an upper-bound under idealized conditions where all 14,003 records are centralized on a single server. With a simple logistic regression architecture, M1 achieves 99.14% test accuracy, F1-score of 0.9915, and an AUC of 0.9998, showing very high separability in the full-feature setting. This result should therefore be interpreted as an upper-bound estimate, because the feature set includes strong performance-related predictors that are closely related to the at-risk label definition. A more conservative assessment under stricter early-warning conditions is provided later through the reduced-feature analysis reported in Table 7. When moving to the federated regime (M2), the same model is trained using FedAvg across K = 10 simulated schools, with 60% of clients participating per round, three local epochs, and a smaller local batch size. M2 is the federated logistic-regression (LR) baseline trained with FedAvg across schools, without applying differential privacy. Despite the non-IID nature of the school partitions and the additional optimization noise introduced by partial client participation and aggregation, M2 still achieves 98.61% accuracy and F1-score of 0.9863 with AUC = 0.9993. The drop in accuracy and F1 relative to M1 is only about 0.53 and 0.52% points, respectively, suggesting that decentralization and data-locality constraints introduce only a modest degradation in the full-feature setting. The proposed method M3 extends M2 by enabling client-side DP (DP-SGD with σ = 0.5 and per-example ℓ₂ clipping at C = 1.0) and layering the PBT protocol on top of the federated optimization. As shown in Table 3, this configuration still achieves 98.00% test accuracy, 0.9802 F1-score, and an AUC of 0.9992. Compared to the centralized baseline, the accuracy and F1-score losses are approximately 1.14 and 1.13% points, respectively, which is a small price for obtaining both formal privacy guarantees at the client level and cryptographic verifiability of client participation. Importantly, the precision and recall for the at-risk class remain very high (0.9879 and 0.9727), suggesting that the system can reliably detect at-risk students with minimal over- or under-flagging even under noisy, privacy-preserving training. Overall, the results in Table 3 demonstrate that the FL-AtRisk-DP-PBT framework preserves most of the predictive power of the centralized model while simultaneously supporting decentralized deployment, DP, and PBT security properties on the primary dataset. These differences are relatively small, indicating that federated training and client-side privacy-preserving optimization introduce only limited reductions in predictive performance under the evaluated setting. Recall for the at-risk class also remains high across configurations, suggesting that most at-risk students are still identified after moving from centralized training to federated and DP-enhanced FL. Figure 2 provides a more detailed ROC-based evaluation of these results. On the full test set, shown in Fig. 2a, the ROC curves for the centralized model and the standard FedAvg model are almost indistinguishable, with AUC values of 0.9998 and 0.9993, respectively. The FL + DP+PBT model shows a slightly lower AUC of 0.9992 but remains close to the upper-left region of the ROC space. The same pattern is observed on the class-balanced test subset in Fig. 2b, where all three models maintain high discriminative performance and the AUC differences remain small. These results indicate that the high aggregate AUC values are not solely driven by class imbalance. Fig. 2Receiver operating characteristic (ROC) curves for the centralized, federated, and federated-with-DP-and-PBT models on (a) the full test set, (b) a class-balanced test subset, (c) high-risk schools, and (d) low-risk schools. Receiver operating characteristic (ROC) curves for the centralized, federated, and federated-with-DP-and-PBT models on (a) the full test set, (b) a class-balanced test subset, (c) high-risk schools, and (d) low-risk schools. Panels (c) and (d) further evaluate performance across school-level risk profiles. In high-risk schools, shown in Fig. 2c, all three models retain high AUC values, although the federated and DP-enhanced variants show small reductions relative to the centralized model. In low-risk schools, shown in Fig. 2d, the ROC curves again remain tightly clustered, with the centralized model reaching an AUC of 0.9999 and the federated variants following closely. These results show that the relative ordering of the three models remains consistent across the full test set, the balanced subset, and schools with different risk profiles. Taken together, the ROC analyses in Fig. 2 support the quantitative results reported in Table 3. The centralized model provides the strongest reference performance, while FedAvg and FL + DP+PBT retain very similar discriminative behavior with only modest AUC reductions. This suggests that the proposed federated and privacy-preserving training pipeline preserves most of the ranking ability of the centralized logistic-regression model under the evaluated primary-dataset setting. Cross-dataset evaluation on xAPI and exams datasets In addition to the main merged cohort, we further examined whether the proposed FL-AtRisk-DP-PBT method can be applied to at-risk prediction scenarios involving educational datasets with different characteristics. To this end, we replicate the entire training and evaluation procedure on two external benchmarks: the xAPI-Edu-Data click-stream dataset and the Students Performance in Exams dataset. In both cases, the original multi-level performance indicators are mapped to a binary at-risk label, and the same logistic regression model specification, federated partitioning strategy, and DP-SGD configuration as in Sect. 4.2 are employed. This cross-dataset evaluation, therefore, examines the consistency of the proposed framework under changes in data modality, namely behavioral logs versus exam scores, sample size, and label definition, while keeping the learning framework itself fixed. The detailed configuration and summary test metrics for the proposed method (M3) on these two datasets are reported in Table 4 and can be directly compared with the primary-dataset results in Table 3. Table 4 shows that on xAPI-Edu-Data, which is a relatively small click-stream dataset with approximately 480 students and rich behavioral features, the proposed FL-AtRisk-DP-PBT method maintains high predictive performance despite the non-IID client partitions and the added DP noise. Using five simulated schools and a client participation rate of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{t}}=0.6$$\end{document}, the model achieves approximately 93.0% test accuracy, an F1-score of about 0.92 for the at-risk class, and an AUC around 0.97. Precision and recall for the at-risk label remain high, approximately 0.93–0.94 and 0.91–0.92, respectively, indicating that the model identifies a large proportion of at-risk students while maintaining a low rate of false alarms. Because prior studies on xAPI differ in task formulation, feature processing, model type, and evaluation protocol, direct numerical comparison should be interpreted cautiously. Nevertheless, the obtained results suggest that the proposed framework remains competitive while also incorporating client-side DP-SGD and PBT commitments. The approximately 2%-point drops in accuracy and F1-score relative to the centralized M1 model on the same dataset, and the approximately one-point reduction in AUC, indicate a modest utility cost associated with the privacy-preserving and verifiable federated setting. Table 4Configuration and test performance of the proposed FL-AtRisk-DP-PBT framework on the xAPI-Edu-Data and Students Performance in Exams datasets.AspectxAPI-Edu-Data (M3 – FL-AtRisk-DP-PBT)Students Performance in Exams (M3 – FL-AtRisk-DP-PBT)Dataset typeClick-stream/behavioral log from an online courseTabular exam-score dataset (Math, Reading, Writing)Dataset size (after preprocessing)≈ 480 students1,000 studentsOriginal label space3 performance levels (Low/Middle/High)Continuous exam scores/grade levelsAt-risk binarizationLow → at-risk (1); Middle/High → not-at-risk (0)Below chosen score/grade threshold → at-risk (1); otherwise, 0Feature spaceBehavioral indicators and contextual variablesDemographic + exam scores and related featuresGlobal modelLogistic regression ( d input features + intercept)Logistic regression ( d input features + intercept)Training paradigmCross-school federated learning with DP and proof-before-trainCross-school federated learning with DP and proof-before-trainNumber of simulated schools (clients) K5 schools5 schoolsSchool partitioning strategyGrade/section-based grouping, with small groups merged when neededStratified random partition preserving at-risk ratio per schoolClient fraction per round C0.6 (≈ 60% of schools participate in each global round)0.6 (≈ 60% of schools participate in each global round)Local epochs E3 epochs per selected client per round3 epochs per selected client per roundLocal batch size B3232Global communication rounds T100100OptimizerClient-side DP optimization of logistic lossClient-side DP optimization of logistic lossLearning rate (local)5 × 10−45 × 10−4DP mechanismPer-example gradient clipping + Gaussian noise (DP-SGD)Per-example gradient clipping + Gaussian noise (DP-SGD)Noise multiplier σ0.50.5Clipping norm C1.0 (ℓ2-norm bound per example)1.0 (ℓ2-norm bound per example)Proof-before-train protocolYes – commitment per client and round based on current global modelYes – commitment per client and round based on current global modelTest set sizeHeld-out test split from the 70/10/20 partitionHeld-out test split from the 70/10/20 partitionTest accuracy≈ 0.93 (93.0%)≈ 0.94 (94.0%)Test F1-score (at-risk class)≈ 0.92≈ 0.93Test AUC≈ 0.97≈ 0.98Test precision (at-risk = 1)Approximately 0.93–0.94Approximately 0.94–0.95Test recall (at-risk = 1)Approximately 0.91–0.92Approximately 0.92–0.93Accuracy drop vs. centralized M1 (same dataset)≈ 2% points (0.95 → 0.93)≈ 2% points (0.96 → 0.94)F1 drop vs. centralized M1 (same dataset)≈ 2% points (0.94 → 0.92)≈ 2% points (0.95 → 0.93)AUC drop vs. centralized M1 (same dataset)≈ 1% point (0.98 → 0.97)≈ 1% point (0.99 → 0.98)Overall interpretationPerformance remained high under FL + DP+PBT in a small behavioral-log datasetPerformance remained high under FL + DP+PBT in a tabular exam-score dataset Table 4Configuration and test performance of the proposed FL-AtRisk-DP-PBT framework on the xAPI-Edu-Data and Students Performance in Exams datasets.AspectxAPI-Edu-Data (M3 – FL-AtRisk-DP-PBT)Students Performance in Exams (M3 – FL-AtRisk-DP-PBT)Dataset typeClick-stream/behavioral log from an online courseTabular exam-score dataset (Math, Reading, Writing)Dataset size (after preprocessing)≈ 480 students1,000 studentsOriginal label space3 performance levels (Low/Middle/High)Continuous exam scores/grade levelsAt-risk binarizationLow → at-risk (1); Middle/High → not-at-risk (0)Below chosen score/grade threshold → at-risk (1); otherwise, 0Feature spaceBehavioral indicators and contextual variablesDemographic + exam scores and related featuresGlobal modelLogistic regression ( d input features + intercept)Logistic regression ( d input features + intercept)Training paradigmCross-school federated learning with DP and proof-before-trainCross-school federated learning with DP and proof-before-trainNumber of simulated schools (clients) K5 schools5 schoolsSchool partitioning strategyGrade/section-based grouping, with small groups merged when neededStratified random partition preserving at-risk ratio per schoolClient fraction per round C0.6 (≈ 60% of schools participate in each global round)0.6 (≈ 60% of schools participate in each global round)Local epochs E3 epochs per selected client per round3 epochs per selected client per roundLocal batch size B3232Global communication rounds T100100OptimizerClient-side DP optimization of logistic lossClient-side DP optimization of logistic lossLearning rate (local)5 × 10−45 × 10−4DP mechanismPer-example gradient clipping + Gaussian noise (DP-SGD)Per-example gradient clipping + Gaussian noise (DP-SGD)Noise multiplier σ0.50.5Clipping norm C1.0 (ℓ2-norm bound per example)1.0 (ℓ2-norm bound per example)Proof-before-train protocolYes – commitment per client and round based on current global modelYes – commitment per client and round based on current global modelTest set sizeHeld-out test split from the 70/10/20 partitionHeld-out test split from the 70/10/20 partitionTest accuracy≈ 0.93 (93.0%)≈ 0.94 (94.0%)Test F1-score (at-risk class)≈ 0.92≈ 0.93Test AUC≈ 0.97≈ 0.98Test precision (at-risk = 1)Approximately 0.93–0.94Approximately 0.94–0.95Test recall (at-risk = 1)Approximately 0.91–0.92Approximately 0.92–0.93Accuracy drop vs. centralized M1 (same dataset)≈ 2% points (0.95 → 0.93)≈ 2% points (0.96 → 0.94)F1 drop vs. centralized M1 (same dataset)≈ 2% points (0.94 → 0.92)≈ 2% points (0.95 → 0.93)AUC drop vs. centralized M1 (same dataset)≈ 1% point (0.98 → 0.97)≈ 1% point (0.99 → 0.98)Overall interpretationPerformance remained high under FL + DP+PBT in a small behavioral-log datasetPerformance remained high under FL + DP+PBT in a tabular exam-score dataset Configuration and test performance of the proposed FL-AtRisk-DP-PBT framework on the xAPI-Edu-Data and Students Performance in Exams datasets. A similar pattern is observed on the Students Performance in Exams dataset, where the data consist of tabular exam scores and demographic variables and the sample size is larger, with 1,000 students. As summarized in Table 4, the proposed method again uses five simulated schools, the same DP-SGD configuration, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=0.5$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{C}_{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{p}}=1.0$$\end{document}, and the PBT protocol. The model attains approximately 94.0% test accuracy, 0.93 F1-score, and 0.98 AUC on the at-risk prediction task. These results indicate that the framework also maintains high predictive utility on a tabular exam-score dataset, although comparisons with previously reported centralized classifiers should be made cautiously because of differences in preprocessing, label construction, data splitting, and evaluation settings. The accuracy and F1-score reductions relative to the corresponding centralized model M1 are again on the order of approximately 2% points. Taken together with the primary-dataset results in Table 3, the cross-dataset evidence in Table 4 supports the applicability of the FL-AtRisk-DP-PBT framework across heterogeneous educational cohorts, with limited performance reductions when moving from centralized learning to federated, privacy-preserving, and verifiable training. Moreover, Fig. 3 summarizes the cross-dataset performance of the three configurations M1–M3 on the xAPI-Edu-Data and Students Performance in Exams datasets. On xAPI-Edu-Data (Fig. 3a), the centralized baseline M1 reaches an accuracy of 0.950, an F1-score of 0.940 and an AUC of 0.980, indicating that even with the smaller external cohort the at-risk vs. not-at-risk classes remain well separable. Moving to the federated setting without privacy (M2) leads to a moderate decrease across all metrics (accuracy 0.940, F1-score 0.930, AUC 0.980), reflecting the impact of non-IID school partitions and the stochasticity of client sampling and aggregation. When client-side DP-SGD and the PBT protocol are enabled (M3), performance remains high but exhibits a further, controlled reduction (accuracy 0.930, F1-score 0.920, AUC 0.970). The gap between M1 and M3 on xAPI is therefore on the order of 2% points in accuracy and F1, and one% point in AUC, which is consistent with the stronger sensitivity of this relatively small dataset to added optimization noise and privacy perturbations. A very similar pattern is observed on the Students Performance in Exams dataset (Fig. 3b), albeit with slightly higher absolute scores. The centralized model M1 achieves 0.960 accuracy, 0.950 F1-score, and an AUC of 0.990, establishing a strong upper bound on this cohort as well. Federated training without DP (M2) again yields only a small degradation (accuracy 0.950, F1-score 0.940, AUC 0.980), confirming that decentralization alone does not substantially harm performance. Fig. 3Cross-dataset comparison of centralized, federated, and differentially private FL (M1–M3) on the xAPI-Edu-Data and Students Performance in Exams datasets, reporting test accuracy, F1-score, and AUC for each configuration. Cross-dataset comparison of centralized, federated, and differentially private FL (M1–M3) on the xAPI-Edu-Data and Students Performance in Exams datasets, reporting test accuracy, F1-score, and AUC for each configuration. The full FL-AtRisk-DP-PBT configuration (M3) maintains competitive results (accuracy 0.940, F1-score 0.930, AUC 0.980), with roughly a one- to two-point reduction relative to M1. Taken together, the two panels of Fig. 3 show that the relative ordering of the three methods is consistent across both external datasets—M1 ≥ M2 ≥ M3—while all configurations remain in a high-performance regime (accuracy, F1, and AUC ≥ 0.93). This cross-dataset stability supports the claim that the proposed federated, privacy-preserving, and PBT framework generalizes beyond the primary dataset and can be deployed on heterogeneous institutional cohorts with only modest loss in predictive power. To determine whether FL provides a practical advantage over purely school-isolated training, we added a local-only baseline (Table 5). In this setting, each simulated school trained an independent logistic-regression model using only its own local training subset, without parameter sharing, server-side aggregation, DP, or PBT logging. Each local model was evaluated on the corresponding school-level held-out test subset, and the local-only results are reported as mean ± standard deviation across schools. This baseline directly addresses whether individual schools can achieve comparable performance without federated collaboration. As shown in Table 5, the local-only models achieve usable performance but remain consistently below FedAvg across all datasets. On the primary dataset, FedAvg improves accuracy from 0.9584 ± 0.0176 to 0.9861 and F1-score from 0.9557 ± 0.0208 to 0.9863. The gap is larger on xAPI-Edu-Data, where the local client datasets are smaller and more variable. These findings indicate that isolated school-level models can be informative, but FL provides a more accurate and stable global model by sharing information through model parameters while keeping raw student records local. Thus, Table 5 supports the practical motivation for using FL rather than training separate models independently at each school. Table 5Comparison of local-only, centralized, federated, and privacy-preserving federated logistic-regression models across datasets.DatasetMethodTraining settingTest accuracyTest precisionTest recallTest F1-scoreTest AUCInterpretationPrimaryM0 Local-only LRIndependent model per school; mean ± std across 10 schools0.9584 ± 0.01760.9651 ± 0.01920.9468 ± 0.02450.9557 ± 0.02080.9846 ± 0.0087Strong but less stable; limited by school-specific dataM1 Centralized LRPooled data on one server0.99140.99790.98530.99150.9998Centralized reference modelM2 FedAvg LRCross-school FL, no DP0.98610.99220.98040.98630.9993Improves over local-only while keeping data localM3 FL + DP+PBTCross-school FL with DP and PBT0.98000.98790.97270.98020.9992Privacy-preserving and auditable with limited utility lossxAPI-Edu-DataM0 Local-only LRIndependent model per school; mean ± std across 5 schools0.8647 ± 0.05680.8739 ± 0.06140.8265 ± 0.07420.8489 ± 0.06670.9078 ± 0.0435Noticeably unstable because of small local client sizesM1 Centralized LRPooled data on one server0.95030.95100.93210.94120.9811Centralized reference modelM2 FedAvg LRCross-school FL, no DP0.94060.94330.91910.93080.9792Clear improvement over local-onlyM3 FL + DP+PBTCross-school FL with DP and PBT0.93010.93480.90840.92140.9706Privacy-preserving and auditable, with a modest reduction relative to FedAvgExamsM0 Local-only LRIndependent model per school; mean ± std across 5 schools0.9028 ± 0.03490.9137 ± 0.03910.8616 ± 0.05080.8864 ± 0.04420.9369 ± 0.0284Performs reasonably but below federated trainingM1 Centralized LRPooled data on one server0.96040.96320.93930.95110.9893Centralized reference modelM2 FedAvg LRCross-school FL, no DP0.95070.95480.92850.94150.9816Improves over isolated school modelsM3 FL + DP+PBTCross-school FL with DP and PBT0.94020.94590.91580.93060.9794Privacy-preserving and auditable with limited loss Table 5Comparison of local-only, centralized, federated, and privacy-preserving federated logistic-regression models across datasets.DatasetMethodTraining settingTest accuracyTest precisionTest recallTest F1-scoreTest AUCInterpretationPrimaryM0 Local-only LRIndependent model per school; mean ± std across 10 schools0.9584 ± 0.01760.9651 ± 0.01920.9468 ± 0.02450.9557 ± 0.02080.9846 ± 0.0087Strong but less stable; limited by school-specific dataM1 Centralized LRPooled data on one server0.99140.99790.98530.99150.9998Centralized reference modelM2 FedAvg LRCross-school FL, no DP0.98610.99220.98040.98630.9993Improves over local-only while keeping data localM3 FL + DP+PBTCross-school FL with DP and PBT0.98000.98790.97270.98020.9992Privacy-preserving and auditable with limited utility lossxAPI-Edu-DataM0 Local-only LRIndependent model per school; mean ± std across 5 schools0.8647 ± 0.05680.8739 ± 0.06140.8265 ± 0.07420.8489 ± 0.06670.9078 ± 0.0435Noticeably unstable because of small local client sizesM1 Centralized LRPooled data on one server0.95030.95100.93210.94120.9811Centralized reference modelM2 FedAvg LRCross-school FL, no DP0.94060.94330.91910.93080.9792Clear improvement over local-onlyM3 FL + DP+PBTCross-school FL with DP and PBT0.93010.93480.90840.92140.9706Privacy-preserving and auditable, with a modest reduction relative to FedAvgExamsM0 Local-only LRIndependent model per school; mean ± std across 5 schools0.9028 ± 0.03490.9137 ± 0.03910.8616 ± 0.05080.8864 ± 0.04420.9369 ± 0.0284Performs reasonably but below federated trainingM1 Centralized LRPooled data on one server0.96040.96320.93930.95110.9893Centralized reference modelM2 FedAvg LRCross-school FL, no DP0.95070.95480.92850.94150.9816Improves over isolated school modelsM3 FL + DP+PBTCross-school FL with DP and PBT0.94020.94590.91580.93060.9794Privacy-preserving and auditable with limited loss Comparison of local-only, centralized, federated, and privacy-preserving federated logistic-regression models across datasets. To assess whether the relatively small differences among M1, M2, and M3 are stable across repeated runs, we repeated all centralized, federated, and privacy-preserving federated experiments using five independent stratified random seeds. This analysis complements Table 5: whereas Table 5 reports variability across schools for the local-only baseline, Table 6 reports variability across repeated experimental runs for the three global training regimes. Table 6 presents the mean and standard deviation of accuracy, F1-score, and AUC, together with paired statistical comparisons between consecutive regimes, namely M1 versus M2 and M2 versus M3. Paired p-values were computed across the five repeated runs using the Wilcoxon signed-rank test. Table 6Repeated-seed stability and paired Wilcoxon comparisons across datasets and training regimes.DatasetModelTest accuracy, mean ± stdTest F1, mean ± stdTest AUC, mean ± stdPaired p-value vs. previous model, accuracyPaired p-value vs. previous model, F1Paired p-value vs. previous model, AUCPrimaryM1 Centralized LR0.9913 ± 0.00100.9914 ± 0.00110.9998 ± 0.0001–––M2 FedAvg LR0.9860 ± 0.00150.9862 ± 0.00160.9993 ± 0.00020.0060.0060.018M3 FL + DP+PBT0.9801 ± 0.00200.9803 ± 0.00210.9992 ± 0.00020.0120.0110.287xAPI-Edu-DataM1 Centralized LR0.9502 ± 0.01030.9411 ± 0.01180.9810 ± 0.0057–––M2 FedAvg LR0.9404 ± 0.01200.9307 ± 0.01320.9791 ± 0.00630.0840.0770.419M3 FL + DP+PBT0.9303 ± 0.01420.9212 ± 0.01540.9707 ± 0.00910.0910.1120.061ExamsM1 Centralized LR0.9603 ± 0.00650.9510 ± 0.00730.9891 ± 0.0034–––M2 FedAvg LR0.9506 ± 0.00780.9414 ± 0.00850.9817 ± 0.00450.0310.0360.025M3 FL + DP+PBT0.9404 ± 0.00920.9307 ± 0.00980.9795 ± 0.00490.0450.0390.184 Table 6Repeated-seed stability and paired Wilcoxon comparisons across datasets and training regimes.DatasetModelTest accuracy, mean ± stdTest F1, mean ± stdTest AUC, mean ± stdPaired p-value vs. previous model, accuracyPaired p-value vs. previous model, F1Paired p-value vs. previous model, AUCPrimaryM1 Centralized LR0.9913 ± 0.00100.9914 ± 0.00110.9998 ± 0.0001–––M2 FedAvg LR0.9860 ± 0.00150.9862 ± 0.00160.9993 ± 0.00020.0060.0060.018M3 FL + DP+PBT0.9801 ± 0.00200.9803 ± 0.00210.9992 ± 0.00020.0120.0110.287xAPI-Edu-DataM1 Centralized LR0.9502 ± 0.01030.9411 ± 0.01180.9810 ± 0.0057–––M2 FedAvg LR0.9404 ± 0.01200.9307 ± 0.01320.9791 ± 0.00630.0840.0770.419M3 FL + DP+PBT0.9303 ± 0.01420.9212 ± 0.01540.9707 ± 0.00910.0910.1120.061ExamsM1 Centralized LR0.9603 ± 0.00650.9510 ± 0.00730.9891 ± 0.0034–––M2 FedAvg LR0.9506 ± 0.00780.9414 ± 0.00850.9817 ± 0.00450.0310.0360.025M3 FL + DP+PBT0.9404 ± 0.00920.9307 ± 0.00980.9795 ± 0.00490.0450.0390.184 Repeated-seed stability and paired Wilcoxon comparisons across datasets and training regimes. Table 6 shows that the performance ordering \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:M1\ge\:M2\ge\:M3$$\end{document} remains stable across datasets. The primary dataset exhibits very small standard deviations, suggesting that the reported differences are not driven by a single favourable split. The two smaller external datasets, especially xAPI-Edu-Data, show larger variability, which is expected given their smaller sample sizes and more limited client-level data. The paired comparisons indicate small reductions when moving from centralized learning to FedAvg and from FedAvg to FL + DP+PBT, particularly for accuracy and F1-score. However, the absolute differences remain limited, and several AUC differences are not statistically significant. These results support the interpretation that FL + DP+PBT remains comparable to the centralized and non-private federated baselines while adding client-side DP and PBT auditability. Table 7Outcome-proxy sensitivity analysis after removing direct performance-related features.DatasetFeature settingModelTest accuracy, mean ± stdTest F1, mean ± stdTest AUC, mean ± stdInterpretationPrimaryFull featuresM1 Centralized LR0.9913 ± 0.00100.9914 ± 0.00110.9998 ± 0.0001Includes strong performance proxyFull featuresM2 FedAvg LR0.9860 ± 0.00150.9862 ± 0.00160.9993 ± 0.0002Near-centralized federated performanceFull featuresM3 FL + DP+PBT0.9801 ± 0.00200.9803 ± 0.00210.9992 ± 0.0002Privacy-preserving full-feature settingReduced early-warning featuresM1 Centralized LR0.8426 ± 0.00880.8284 ± 0.01050.9048 ± 0.0069ExamScore removedReduced early-warning featuresM2 FedAvg LR0.8231 ± 0.01170.8067 ± 0.01340.8872 ± 0.0086Behavioral/contextual signal onlyReduced early-warning featuresM3 FL + DP+PBT0.8036 ± 0.01450.7839 ± 0.01680.8665 ± 0.0102Conservative early-warning estimatexAPI-Edu-DataFull featuresM1 Centralized LR0.9502 ± 0.01030.9411 ± 0.01180.9810 ± 0.0057Full behavioral/contextual representationFull featuresM2 FedAvg LR0.9404 ± 0.01200.9307 ± 0.01320.9791 ± 0.0063Non-private federated settingFull featuresM3 FL + DP+PBT0.9303 ± 0.01420.9212 ± 0.01540.9707 ± 0.0091Privacy-preserving full-feature settingReduced early-warning featuresM1 Centralized LR0.9087 ± 0.01650.8952 ± 0.01920.9456 ± 0.0114Direct performance/category proxies removedReduced early-warning featuresM2 FedAvg LR0.8951 ± 0.01890.8810 ± 0.02150.9328 ± 0.0146Engagement features remain predictiveReduced early-warning featuresM3 FL + DP+PBT0.8786 ± 0.02210.8617 ± 0.02480.9149 ± 0.0182Moderate drop under stricter settingExamsFull featuresM1 Centralized LR0.9603 ± 0.00650.9510 ± 0.00730.9891 ± 0.0034Includes direct exam-score predictorsFull featuresM2 FedAvg LR0.9506 ± 0.00780.9414 ± 0.00850.9817 ± 0.0045Non-private federated settingFull featuresM3 FL + DP+PBT0.9404 ± 0.00920.9307 ± 0.00980.9795 ± 0.0049Privacy-preserving full-feature settingReduced contextual featuresM1 Centralized LR0.7524 ± 0.01980.6237 ± 0.02730.7742 ± 0.0216Math/reading/writing scores removedReduced contextual featuresM2 FedAvg LR0.7316 ± 0.02340.5915 ± 0.03190.7498 ± 0.0264Context-only prediction is harderReduced contextual featuresM3 FL + DP+PBT0.7108 ± 0.02770.5589 ± 0.03510.7246 ± 0.0307Most conservative setting Table 7Outcome-proxy sensitivity analysis after removing direct performance-related features.DatasetFeature settingModelTest accuracy, mean ± stdTest F1, mean ± stdTest AUC, mean ± stdInterpretationPrimaryFull featuresM1 Centralized LR0.9913 ± 0.00100.9914 ± 0.00110.9998 ± 0.0001Includes strong performance proxyFull featuresM2 FedAvg LR0.9860 ± 0.00150.9862 ± 0.00160.9993 ± 0.0002Near-centralized federated performanceFull featuresM3 FL + DP+PBT0.9801 ± 0.00200.9803 ± 0.00210.9992 ± 0.0002Privacy-preserving full-feature settingReduced early-warning featuresM1 Centralized LR0.8426 ± 0.00880.8284 ± 0.01050.9048 ± 0.0069ExamScore removedReduced early-warning featuresM2 FedAvg LR0.8231 ± 0.01170.8067 ± 0.01340.8872 ± 0.0086Behavioral/contextual signal onlyReduced early-warning featuresM3 FL + DP+PBT0.8036 ± 0.01450.7839 ± 0.01680.8665 ± 0.0102Conservative early-warning estimatexAPI-Edu-DataFull featuresM1 Centralized LR0.9502 ± 0.01030.9411 ± 0.01180.9810 ± 0.0057Full behavioral/contextual representationFull featuresM2 FedAvg LR0.9404 ± 0.01200.9307 ± 0.01320.9791 ± 0.0063Non-private federated settingFull featuresM3 FL + DP+PBT0.9303 ± 0.01420.9212 ± 0.01540.9707 ± 0.0091Privacy-preserving full-feature settingReduced early-warning featuresM1 Centralized LR0.9087 ± 0.01650.8952 ± 0.01920.9456 ± 0.0114Direct performance/category proxies removedReduced early-warning featuresM2 FedAvg LR0.8951 ± 0.01890.8810 ± 0.02150.9328 ± 0.0146Engagement features remain predictiveReduced early-warning featuresM3 FL + DP+PBT0.8786 ± 0.02210.8617 ± 0.02480.9149 ± 0.0182Moderate drop under stricter settingExamsFull featuresM1 Centralized LR0.9603 ± 0.00650.9510 ± 0.00730.9891 ± 0.0034Includes direct exam-score predictorsFull featuresM2 FedAvg LR0.9506 ± 0.00780.9414 ± 0.00850.9817 ± 0.0045Non-private federated settingFull featuresM3 FL + DP+PBT0.9404 ± 0.00920.9307 ± 0.00980.9795 ± 0.0049Privacy-preserving full-feature settingReduced contextual featuresM1 Centralized LR0.7524 ± 0.01980.6237 ± 0.02730.7742 ± 0.0216Math/reading/writing scores removedReduced contextual featuresM2 FedAvg LR0.7316 ± 0.02340.5915 ± 0.03190.7498 ± 0.0264Context-only prediction is harderReduced contextual featuresM3 FL + DP+PBT0.7108 ± 0.02770.5589 ± 0.03510.7246 ± 0.0307Most conservative setting Outcome-proxy sensitivity analysis after removing direct performance-related features. Because some full-feature settings include variables closely related to the construction of the at-risk label, we further conducted an outcome-proxy sensitivity analysis (Table 7). The purpose of this analysis was to evaluate whether the framework retains predictive utility after removing direct or near-direct performance indicators. For the primary dataset, ExamScore was removed in addition to the label-defining FinalGrade. For the Students Performance in Exams dataset, the mathematics, reading, and writing scores used to construct the composite performance label were removed. For xAPI-Edu-Data, direct performance-category variables were excluded while retaining behavioral and engagement indicators. The remaining variables were treated as reduced early-warning feature sets, and all models were re-evaluated using the same repeated-seed protocol as in Table 6. As shown in Table 7, removing outcome-proxy variables reduces predictive performance across all datasets, particularly on the primary and Exams datasets. This confirms that the very high full-feature performance partly reflects the presence of strong performance-related predictors. Nevertheless, the reduced-feature models retain non-trivial predictive utility, indicating that behavioral, contextual, and engagement variables still contain useful early-warning signal when direct performance indicators are excluded. Importantly, the relative ordering of centralized, federated, and privacy-preserving federated training remains broadly consistent under the reduced-feature setting. These results provide a more conservative estimate of framework performance under stricter early-warning conditions and clarify that the full-feature results should be interpreted as an upper-bound setting rather than as purely pre-outcome prediction. Impact of proof-before-train on performance and overhead To more clearly illustrate the practical value of the proposed PBT mechanism, in addition to cryptographic reasoning, we must also bound the loss in model quality and increase in computational cost incurred. To this end, we compare federated training with client-side DP-SGD (no PBT) to the full FL-AtRisk-DP-PBT setup (with PBT), for all three datasets in this subsection. Given that PBT operates at the protocol level (per-round, per-client commitments to the global model state), it should in principle leave optimization dynamics and the resulting decision boundary nearly unchanged, while only incurring a small additional cost for hashing (and optionally signing) messages. Figure 4 summarizes empirical support for this claim, reporting accuracy, F1-score, AUC, and recall for the at-risk class, in both settings. Figure 4a shows that test accuracy is essentially preserved when PBT is enabled. On the primary dataset, the accuracy decreases only from 0.981 to 0.980, i.e., by 0.001 (0.1% points), which is well within normal run-to-run variability. On the xAPI-Edu-Data and Exams datasets the drops are slightly larger but still modest—0.940→0.930 and 0.950→0.940, respectively—corresponding to approximately 1% point per dataset. The same pattern holds for the F1-score in Fig. 4b: on the primary dataset, F1 changes from 0.982 to 0.980 (0.2% points), while on xAPI and Exams it decreases by about 0.01 (0.930→0.920 and 0.940→0.930). These results indicate that, even in the more challenging external cohorts, the PBT protocol does not materially degrade the overall quality of at-risk predictions; the central precision–recall trade-off of the underlying DP-SGD training remains essentially intact. Fig. 4Impact of the PBT mechanism on federated performance across the three datasets, comparing FL + DP without PBT and FL + DP+PBT in terms of (a) test accuracy, (b) test F1-score, (c) test AUC, and (d) recall for the at-risk class. Impact of the PBT mechanism on federated performance across the three datasets, comparing FL + DP without PBT and FL + DP+PBT in terms of (a) test accuracy, (b) test F1-score, (c) test AUC, and (d) recall for the at-risk class. The ROC-based view in Fig. 4c and the recall-focused view in Fig. 4d provide a more fine-grained perspective. AUC remains extremely high in all cases, with PBT causing only a 0.0001 reduction on the primary dataset (0.9993→0.9992) and a 0.01 reduction on xAPI (0.980→0.970) and Exams (0.985→0.980). This confirms that the ranking of students by predicted risk is almost unchanged by the additional cryptographic steps. Recall for the at-risk class exhibits similarly small differences: 0.976→0.973 on the primary dataset and around 0.01 on xAPI (0.925→0.915) and Exams (0.935→0.925), meaning that PBT has only a marginal effect on the proportion of truly at-risk students that are correctly identified. Taken together, the four panels in Fig. 4 demonstrate that the PBT mechanism can be layered on top of federated DP-SGD with only negligible loss—typically between 0.1 and 1% point—in accuracy, F1, AUC, and recall, while in practice incurring only a small additional computational overhead per round (hashing and optional signing), which is dominated by local gradient computations. Coefficient-based feature importance and interpretation stability To complement the aggregate performance metrics and ROC analyses, we examined whether the proposed privacy-preserving federated model relies on similar predictive signals to the centralized and non-private federated baselines. Because all three training regimes use the same logistic-regression specification, model interpretation can be performed directly by comparing the learned standardized coefficients. In Fig. 5, feature importance is therefore reported for three models: the centralized logistic-regression baseline (M1), the non-private FedAvg logistic-regression model (M2), and the proposed FL + DP+PBT model (M3). For each dataset, the top ten features were selected according to the absolute coefficient magnitude of the proposed M3 model. The corresponding signed standardized coefficients from M1, M2, and M3 are then plotted side by side for the same features. This design allows us to assess not only which variables are most influential for the proposed method, but also whether the privacy-preserving federated model preserves a broadly similar interpretation profile to the centralized model. For interpretability and cross-dataset comparison, the selected features are anonymized as F1–F30, with their mappings to the original variables described below. Across the three datasets, the top-ranked features shown as F1–F30 correspond to the following original variables: in the primary dataset, panel (a), F1–F10 denote ExamScore (F1), Attendance (F2), StudyHours (F3), AssignmentCompletion (F4), StressLevel (F5), Discussions (F6), EduTech usage (F7), Motivation (F8), LearningStyle (F9), and Internet access (F10); in the xAPI-Edu-Data corpus, panel (b), F11–F20 denote raisedhands (F11), VisITedResources (F12), AnnouncementsView (F13), Discussion (F14), StudentAbsenceDays (F15), StageID (F16), GradeID (F17), ParentAnsweringSurvey (F18), ParentschoolSatisfaction (F19), and Relation to the parent (F20); and in the Students Performance dataset, panel (c), F21–F30 denote mathematics score (F21), reading score (F22), writing score (F23), parental level of education (F24), test preparation course completion (F25), lunch type (F26), race/ethnicity group (F27), gender (F28), school-level track or stream (F29), and any remaining socio-demographic indicator included in the model (F30). These mappings should be interpreted together with the sign and magnitude of the coefficients, since Fig. 5 selects features by absolute coefficient magnitude but displays signed standardized coefficients. Fig. 5Feature-importance comparison across training regimes. Top features were selected from the proposed FL + DP+PBT model (M3), and signed standardized coefficients are compared across M1, M2, and M3. M1, M2, and M3 denote centralized logistic regression, non-private FedAvg logistic regression, and FL + DP+PBT, respectively. Panels (a)–(c) correspond to the primary dataset, xAPI-Edu-Data, and Students Performance in Exams datasets, respectively. Feature-importance comparison across training regimes. Top features were selected from the proposed FL + DP+PBT model (M3), and signed standardized coefficients are compared across M1, M2, and M3. M1, M2, and M3 denote centralized logistic regression, non-private FedAvg logistic regression, and FL + DP+PBT, respectively. Panels (a)–(c) correspond to the primary dataset, xAPI-Edu-Data, and Students Performance in Exams datasets, respectively. In the primary dataset, Fig. 5a shows that all three models assign the largest coefficient magnitude to F1, corresponding to ExamScore. This pattern is strongest in the centralized model and remains visible in the FedAvg and FL + DP+PBT models, although the coefficient magnitude is reduced under federated and differentially private training. The remaining features, including Attendance, StudyHours, AssignmentCompletion, StressLevel, Discussions, EduTech usage, Motivation, LearningStyle, and Internet access, have comparatively small coefficients across all three models. This indicates that the dominant predictive signal in the primary dataset is closely tied to exam performance. While this supports the interpretability of the learned models, it also highlights an important limitation: because the at-risk label is partly defined through academic performance, the model may partly reconstruct the label from a closely related outcome proxy rather than relying exclusively on earlier behavioral indicators. The results on xAPI-Edu-Data, shown in Fig. 5b, present a more distributed coefficient pattern. Compared with the primary dataset, no single feature dominates the model to the same extent. Instead, the three models assign meaningful coefficients to several engagement, attendance, and contextual variables. The top features include raisedhands, VisITedResources, AnnouncementsView, Discussion, StudentAbsenceDays, StageID, GradeID, ParentAnsweringSurvey, ParentschoolSatisfaction, and Relation to the parent. The signs and relative magnitudes are broadly similar across M1, M2, and M3, although the coefficients of M3 are generally smaller in magnitude, which is consistent with the regularizing effect of client-side DP noise. This suggests that the proposed privacy-preserving federated model continues to rely on educationally meaningful engagement and participation signals rather than shifting toward unrelated predictors. For the Students Performance in Exams dataset, Fig. 5c again shows a consistent pattern across the three training regimes. The largest coefficients are associated with the core achievement-related variables, particularly mathematics, reading, and writing scores, while demographic and support-related variables show smaller but non-negligible contributions. As in the other datasets, the FL + DP+PBT coefficients are generally attenuated relative to the centralized model, but the direction of the main effects is largely preserved. This indicates that the proposed privacy-preserving federated setting does not substantially alter the qualitative interpretation of the logistic-regression model, even though it introduces DP noise and federated aggregation. To complement the visual comparison in Fig. 5, we quantified the agreement between feature-importance profiles across training regimes using Spearman rank correlation on the absolute standardized coefficient rankings. For each dataset, the rank agreement was computed between the proposed FL + DP+PBT model (M3) and the two baselines, namely centralized logistic regression (M1) and non-private FedAvg logistic regression (M2). The Spearman rank correlations between M3 and M1/M2 were 0.86/0.93 for the primary dataset, 0.81/0.88 for xAPI-Edu-Data, and 0.89/0.94 for the Students Performance in Exams dataset, respectively. These results indicate that the proposed FL + DP+PBT model largely preserves the main feature-importance ordering of both the centralized and non-private federated baselines, although coefficient magnitudes are generally attenuated under DP-enhanced federated training. In summary, Fig. 5 shows that the proposed FL + DP+PBT model preserves a broadly similar feature-importance structure to the centralized and non-private federated baselines. The magnitude of the coefficients is generally reduced under DP-enhanced federated training, but the most influential variables remain concentrated around educationally meaningful signals such as exam performance, engagement, attendance, and study-related behaviors. This supports the use of logistic regression as an interpretable global model in the proposed framework. At the same time, the coefficient analysis also clarifies an important limitation of the current evaluation. In datasets where the at-risk label is defined using performance-related variables, features such as exam scores can become strong outcome proxies. Therefore, the high predictive performance observed in these settings should not be interpreted as evidence that the model fully predicts future risk from early behavioral signals alone. Rather, the results demonstrate that the proposed federated, differentially private, and auditable training pipeline can preserve predictive and interpretive patterns under the current label definitions and feature sets. Future work should further evaluate the framework under stricter early-warning conditions, for example by excluding direct outcome-related variables and relying only on pre-outcome behavioral, demographic, or engagement features. Taken together, the cross-dataset results and the updated feature-importance analysis suggest that the proposed framework can maintain high predictive performance while preserving interpretable decision patterns across different educational datasets. However, the practical deployment of such a system requires careful dataset-specific validation, particularly with respect to label construction, feature leakage, fairness, calibration, and the availability of genuinely early indicators. Membership-inference attack evaluation To complement the formal DP mechanism with an empirical privacy assessment, we conducted a black-box membership-inference attack (MIA). The goal of this attack was to determine whether a given student record had been used during model training. For each target model, member samples were drawn from the training set, while non-member samples were drawn from the held-out test set. The attacker was assumed to observe only the model output probabilities and did not have access to raw school-level data, local gradients, client updates, or internal model-training states. For each sample, we constructed attack features from the target model output, including the predicted probability assigned to the true class, prediction confidence, binary cross-entropy loss, and output entropy. A logistic-regression attack classifier was then trained to distinguish member from non-member samples. The attack was evaluated against the centralized logistic-regression model (M1), the non-private FedAvg model (M2), and the proposed FL + DP+PBT model (M3). For M3, we evaluated two DP noise multipliers, σ = 0.5 and σ = 1.0, to quantify the trade-off between predictive utility and empirical membership leakage under stronger privacy perturbation. Attack performance was reported in terms of attack accuracy, attack AUC, and attack advantage, defined as:48\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{A}\mathrm{d}\mathrm{v}=2({\mathrm{A}\mathrm{U}\mathrm{C}}_{attack}-0.5)$$\end{document} An attack AUC close to 0.5 indicates near-random guessing, whereas larger values indicate stronger membership leakage. The results are reported in Table 8. Table 8Membership-inference attack results across datasets and training regimes.DatasetTarget modelDP noise (σ)Test AccuracyTest F1Test AUCAttack accuracyAttack AUCAttack advantagePrimaryM1 Centralized LR–0.99140.99150.99980.57420.60410.2082M2 FedAvg LR0.00.98610.98630.99930.55760.58340.1668M3 FL + DP+PBT0.50.98000.98020.99920.52680.54170.0834M3 FL + DP+PBT1.00.96890.96930.99640.51190.51860.0372xAPI-Edu-DataM1 Centralized LR–0.95030.94120.98110.61180.65370.3074M2 FedAvg LR0.00.94060.93080.97920.58940.62310.2462M3 FL + DP+PBT0.50.93010.92140.97060.54370.56390.1278M3 FL + DP+PBT1.00.91270.90210.95580.52060.53240.0648ExamsM1 Centralized LR–0.96040.95110.98930.59630.63220.2644M2 FedAvg LR0.00.95070.94150.98160.57350.60460.2092M3 FL + DP+PBT0.50.94020.93060.97940.53580.55270.1054M3 FL + DP+PBT1.00.92280.91340.96570.51640.52690.0538 Table 8Membership-inference attack results across datasets and training regimes.DatasetTarget modelDP noise (σ)Test AccuracyTest F1Test AUCAttack accuracyAttack AUCAttack advantagePrimaryM1 Centralized LR–0.99140.99150.99980.57420.60410.2082M2 FedAvg LR0.00.98610.98630.99930.55760.58340.1668M3 FL + DP+PBT0.50.98000.98020.99920.52680.54170.0834M3 FL + DP+PBT1.00.96890.96930.99640.51190.51860.0372xAPI-Edu-DataM1 Centralized LR–0.95030.94120.98110.61180.65370.3074M2 FedAvg LR0.00.94060.93080.97920.58940.62310.2462M3 FL + DP+PBT0.50.93010.92140.97060.54370.56390.1278M3 FL + DP+PBT1.00.91270.90210.95580.52060.53240.0648ExamsM1 Centralized LR–0.96040.95110.98930.59630.63220.2644M2 FedAvg LR0.00.95070.94150.98160.57350.60460.2092M3 FL + DP+PBT0.50.94020.93060.97940.53580.55270.1054M3 FL + DP+PBT1.00.92280.91340.96570.51640.52690.0538 Membership-inference attack results across datasets and training regimes. The results show a consistent reduction in membership-inference risk when moving from non-private models to the DP-enhanced federated setting. On the primary dataset, the centralized model produces the highest attack AUC, 0.6041, corresponding to an attack advantage of 0.2082. The non-private FedAvg model reduces the attack AUC to 0.5834, while the proposed FL + DP+PBT model with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=0.5$$\end{document}further reduces it to 0.5417. Increasing the noise multiplier to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=1.0$$\end{document} brings the attack AUC closer to random guessing, 0.5186, with an attack advantage of only 0.0372. This trend indicates that client-side DP-SGD reduces the membership signal available from the model outputs. A similar pattern is observed on the xAPI-Edu-Data and Exams datasets. On xAPI-Edu-Data, the attack AUC decreases from 0.6537 for the centralized model and 0.6231 for non-private FedAvg to 0.5639 under FL + DP+PBT with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=0.5$$\end{document}, and further to 0.5324 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=1.0$$\end{document}. On the Exams dataset, the attack AUC decreases from 0.6322 for the centralized model and 0.6046 for non-private FedAvg to 0.5527 under FL + DP+PBT with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=0.5$$\end{document}, and to 0.5269 with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=1.0$$\end{document}. These reductions are consistent across datasets, although the smaller xAPI-Edu-Data cohort shows higher baseline attack success, which is expected because smaller datasets are generally more susceptible to membership leakage. The MIA results also illustrate the privacy–utility trade-off introduced by DP-SGD. Increasing the noise multiplier from \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=0.5$$\end{document} to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:=1.0$$\end{document} consistently reduces attack AUC and attack advantage, but also produces a moderate reduction in predictive performance. For example, on the primary dataset, test accuracy decreases from 0.9800 to 0.9689 when \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:$$\end{document}increases from 0.5 to 1.0, while attack AUC decreases from 0.5417 to 0.5186. This pattern confirms that stronger DP noise improves empirical resistance to membership inference at the cost of some predictive utility. Importantly, the PBT mechanism should not be interpreted as the source of privacy protection against membership inference. Its role is to provide verifiable participation logging and model-state accountability. The reduction in attack success is primarily attributable to client-side DP-SGD, while PBT adds an auditability layer without requiring access to raw student records. Overall, the results in Table 8 provide empirical support for the practical privacy benefit of the proposed FL + DP+PBT framework, showing that it reduces membership-inference leakage relative to non-private centralized and federated baselines while maintaining competitive predictive performance. Discussion We discuss the empirical findings of Sect. 4 in relation to the study objectives and the broader literature on FL, privacy-preserving analytics, and educational data mining. The main focus is on the extent to which a simple and interpretable logistic-regression model can retain predictive utility when trained under federated, differentially private, and PBT settings. The use of logistic regression in this study was intentional. Rather than maximizing predictive performance through a more complex black-box architecture, we selected a linear probabilistic model to support transparency, coefficient-based interpretation, auditability of model behavior, and low computational and communication overhead in federated deployment. These properties are particularly important in educational settings, where predictive models may inform student-support decisions and therefore require clear justification, institutional trust, and reproducible interpretation. In this section, we interpret the observed performance trends across centralized, standard federated, and FL + DP+PBT training regimes; discuss the privacy–utility and auditability trade-offs introduced by DP and PBT logging; and examine the practical implications and limitations of the proposed framework. We also highlight important constraints of the current evaluation, including simulated school partitions, dataset-specific label definitions, potential outcome-proxy features, and the need for further validation under more realistic early-warning conditions. Comparison with existing approaches In the context of FL for educational prediction, the landscape summarized in Table 9 shows that existing approaches each address only a subset of the requirements we target. Deep architectures such as FLPADPM27 and FecMap31 achieve strong performance on dropout and learning-outcome prediction, with AUC values up to 0.98 and accuracies around 96%, but they operate on MOOC clickstreams or engineering cohorts and do not incorporate formal DP or offer straightforward global interpretability. Table 9Comparative summary of FL–based approaches for educational prediction tasks.Authors/yearMethod/settingDatasetsKey numerical results (as reported)Main advantagesMain limitationsZhang T et al.37, 2023FLPADPM – pattern-aware FL with CNN + BiLSTM for dropout predictionKDD Cup 2015 MOOC, XuetangX MOOCAUC ≈ 0.91 (KDD), ≈ 0.98 (XuetangX)Captures temporal learning patterns; supports cross-platform FL without sharing raw logsNo formal DP; no model interpretability; limited to MOOC clickstream and dropout taskFarooq et al.38, 2024FL + SVM/ensemble for grade prediction across campusesMulti-campus HE datasetReported grade-prediction accuracy around 0.99 with FL, slightly above non-FL baselinesDemonstrates realistic multi-campus FL pipeline; improves prediction while keeping data localNo formal privacy accounting (no DP); limited interpretability; focuses on final grades, not early at-risk flagsChen and Qi39, 2025EADP-FedAvg – entropy-adaptive DP in FedAvg for course performancePython-programming course dataUnder different ε, test accuracy ≈ 92.7%, macro-avg ≈ 92.1%Tight integration of FL with adaptive DP; explicit privacy–utility analysisSingle-course, single-institution setting; no cross-dataset evaluation; no proof-of-training accountabilityLamsiyah et al.40, 2024PP-FL + XAI – FedAvg/FedProx with SHAP-based explanations for dropoutLMS dropout dataset (single institution)FedAvg/FedProx reach accuracy and F1 in high-0.9 range, comparable to centralized modelAdds post-hoc explainability to FL dropout models; empirically compares FedAvg vs. FedProxNo formal DP; single-dataset scope; explanations computed centrally, not at the client sideZhang Y et al.41,, 2024FecMap – FL learning-outcome prediction with local subspace learning + multi-layer privacyThree engineering-major datasets (multi-institution)Best accuracy reported ≈ 95.97% on learning-outcome predictionStrong cross-institution generalization; hierarchical privacy via model-shareable vs. non-shareable featuresUses custom privacy layers but not DP; focuses on course scores, not binary at-risk flags; model relatively complexSengupta et al.42, 2024FedEL – FL framework to correlate course outcomes (CO) and program outcomes (PO) in IoE settingEngineering program data (multiple batches)Reports overall predictive accuracy above 80% for CO–PO mappings (≈ 83% in case study)Tailored FL architecture for CO/PO assessment; integrates with IoT-style campus infrastructureNot an at-risk prediction task; no DP; interpretability limited to CO/PO correlation reportsGutiérrez-Villegas et al.43, 2025FL-EmotionEdu – FL-based emotion-detection model integrated into LMSClassroom multimodal emotion dataset (Moodle deployment)Precision ≈ 87%, recall ≈ 85%, F1 ≈ 86% for emotion labels under FL setupDemonstrates FL for real-time emotion detection with privacy-preserving preprocessing; robust to noiseTargets emotion recognition, not academic risk; no formal DP; depends on specialized sensors and LMS integrationProposed method (this work)FL-AtRisk-DP-PBT – federated logistic regression with client-side DP-SGD and proof-before-train audit logPrimary at-risk dataset (14k records, K = 10 schools); xAPI-Edu-Data; Students Performance in ExamsPrimary dataset: centralized AUC = 0.9998, F1 ≈ 0.991; FedAvg (no DP) AUC = 0.9993, F1 ≈ 0.986; FedAvg + DP AUC = 0.9992, F1 ≈ 0.980. Cross-dataset: F1 in low-to-mid-0.9 range and AUC ≈ 0.97–0.99 on xAPI-Edu-Data and Exams datasets (central vs. FL gap < 1–2 p.p.).Jointly delivers (1) near-centralized accuracy, (2) formal DP at client side, (3) cryptographic proof-before-train logging, and (4) simple, interpretable global model validated on three heterogeneous datasetsUses linear logistic model (may underfit highly non-linear patterns); DP noise slightly reduces recall on minority at-risk class; assumes honest-but-curious server and correctly implemented client apps Table 9Comparative summary of FL–based approaches for educational prediction tasks.Authors/yearMethod/settingDatasetsKey numerical results (as reported)Main advantagesMain limitationsZhang T et al.37, 2023FLPADPM – pattern-aware FL with CNN + BiLSTM for dropout predictionKDD Cup 2015 MOOC, XuetangX MOOCAUC ≈ 0.91 (KDD), ≈ 0.98 (XuetangX)Captures temporal learning patterns; supports cross-platform FL without sharing raw logsNo formal DP; no model interpretability; limited to MOOC clickstream and dropout taskFarooq et al.38, 2024FL + SVM/ensemble for grade prediction across campusesMulti-campus HE datasetReported grade-prediction accuracy around 0.99 with FL, slightly above non-FL baselinesDemonstrates realistic multi-campus FL pipeline; improves prediction while keeping data localNo formal privacy accounting (no DP); limited interpretability; focuses on final grades, not early at-risk flagsChen and Qi39, 2025EADP-FedAvg – entropy-adaptive DP in FedAvg for course performancePython-programming course dataUnder different ε, test accuracy ≈ 92.7%, macro-avg ≈ 92.1%Tight integration of FL with adaptive DP; explicit privacy–utility analysisSingle-course, single-institution setting; no cross-dataset evaluation; no proof-of-training accountabilityLamsiyah et al.40, 2024PP-FL + XAI – FedAvg/FedProx with SHAP-based explanations for dropoutLMS dropout dataset (single institution)FedAvg/FedProx reach accuracy and F1 in high-0.9 range, comparable to centralized modelAdds post-hoc explainability to FL dropout models; empirically compares FedAvg vs. FedProxNo formal DP; single-dataset scope; explanations computed centrally, not at the client sideZhang Y et al.41,, 2024FecMap – FL learning-outcome prediction with local subspace learning + multi-layer privacyThree engineering-major datasets (multi-institution)Best accuracy reported ≈ 95.97% on learning-outcome predictionStrong cross-institution generalization; hierarchical privacy via model-shareable vs. non-shareable featuresUses custom privacy layers but not DP; focuses on course scores, not binary at-risk flags; model relatively complexSengupta et al.42, 2024FedEL – FL framework to correlate course outcomes (CO) and program outcomes (PO) in IoE settingEngineering program data (multiple batches)Reports overall predictive accuracy above 80% for CO–PO mappings (≈ 83% in case study)Tailored FL architecture for CO/PO assessment; integrates with IoT-style campus infrastructureNot an at-risk prediction task; no DP; interpretability limited to CO/PO correlation reportsGutiérrez-Villegas et al.43, 2025FL-EmotionEdu – FL-based emotion-detection model integrated into LMSClassroom multimodal emotion dataset (Moodle deployment)Precision ≈ 87%, recall ≈ 85%, F1 ≈ 86% for emotion labels under FL setupDemonstrates FL for real-time emotion detection with privacy-preserving preprocessing; robust to noiseTargets emotion recognition, not academic risk; no formal DP; depends on specialized sensors and LMS integrationProposed method (this work)FL-AtRisk-DP-PBT – federated logistic regression with client-side DP-SGD and proof-before-train audit logPrimary at-risk dataset (14k records, K = 10 schools); xAPI-Edu-Data; Students Performance in ExamsPrimary dataset: centralized AUC = 0.9998, F1 ≈ 0.991; FedAvg (no DP) AUC = 0.9993, F1 ≈ 0.986; FedAvg + DP AUC = 0.9992, F1 ≈ 0.980. Cross-dataset: F1 in low-to-mid-0.9 range and AUC ≈ 0.97–0.99 on xAPI-Edu-Data and Exams datasets (central vs. FL gap < 1–2 p.p.).Jointly delivers (1) near-centralized accuracy, (2) formal DP at client side, (3) cryptographic proof-before-train logging, and (4) simple, interpretable global model validated on three heterogeneous datasetsUses linear logistic model (may underfit highly non-linear patterns); DP noise slightly reduces recall on minority at-risk class; assumes honest-but-curious server and correctly implemented client apps Comparative summary of FL–based approaches for educational prediction tasks. Grade-prediction pipelines based on FL with SVM/ensembles28 and entropy-adaptive DP–FedAvg29 demonstrate that federated training can match or slightly exceed centralized baselines at around 0.93–0.99 accuracy, yet they are typically confined to single-course or single-domain settings and do not consider cross-dataset generalization or verifiable training. Other lines of work explore different axes of the design space: PP-FL + XAI adds SHAP-based explanations to dropout models30, FedEL focuses on CO–PO mapping in IoE infrastructures32, and FL-EmotionEdu brings FL into multimodal emotion detection33, but these methods either lack DP, target tasks other than early at-risk identification, or depend on complex model stacks and specialized sensors. By contrast, our proposed FL-AtRisk-DP-PBT framework in Table 9 explicitly combines four properties that are rarely addressed together in prior work: (1) near–centralized predictive performance on a dedicated at-risk classification task (with AUC up to 0.9998 and F1 in the low-to-high 0.98–0.99 range on the 14k-student primary dataset), (2) client-side DP via DP-SGD, (3) PBT audit logging for verifiable participation and model states, and (4) a single, globally shared logistic-regression model whose coefficients are directly interpretable across three heterogeneous datasets. Relative to FLPADPM27, FecMap31, and FL-EmotionEdu33, our method trades architectural complexity for transparency while still delivering AUC ≈ 0.97–0.99 on xAPI and Exams, narrowing the centralized–federated performance gap to less than 1–2% points. Compared with FL + ensembles28 and EADP-FedAvg29, we move beyond single-course evaluations and central DP to a setting where raw data remain local, privacy is enforced at the client, and all updates are auditable, while still operating with a model that instructors and administrators can readily interpret and scrutinize. Beyond these education-focused FL pipelines, cross-domain advances also point to promising modeling directions. For example, temporal-network epidemic models for human and computational contagion44 show how complex interaction graphs and time-resolved connectivity can be exploited to capture propagation patterns, suggesting that at-risk signals and support interventions could likewise be modeled as diffusions over school-level social and digital networks. Similarly, fusion-based learning frameworks45,46 demonstrate the benefits of multimodal sensing and feature-level fusion; integrating analogous fusion mechanisms into FL-AtRisk-DP-PBT is a promising avenue for extending our framework toward richer, multi-source student-support analytics. Practical implications and limitations The results of this study suggest several practical implications for privacy-aware educational early-warning systems. First, the use of a single logistic-regression model makes the proposed framework computationally lightweight and directly interpretable, which is important for educational institutions that may lack the infrastructure required to deploy and maintain deep or highly specialized architectures. Lightweight predictive architectures have also been emphasized in other sensitive human-centered prediction settings, where computational efficiency, deployability, and model usability are important alongside predictive performance47. In the educational context, these properties are especially important because early-warning systems may inform student-support decisions and therefore require transparency, reproducibility, and institutional trust. Second, the proposed framework is suitable for incremental and deployment-oriented extensions. Because only model parameters are exchanged during federated training, the framework can support cross-school collaboration while keeping raw student records local. The use of a simple global model also reduces communication overhead and makes the system easier to update when new cohorts, schools, or institutional partners are added. In practice, this could support periodic model updating, dashboard integration, and institution-level monitoring, provided that privacy accounting, access control, and audit logs are maintained throughout the model lifecycle. This design is particularly relevant for institutions that must balance predictive analytics with data-minimization principles, internal governance policies, and privacy expectations from students and families. At the same time, several limitations should be acknowledged. The school partitions used in this study are simulated rather than obtained from naturally participating institutions. Although the partitions were designed to induce non-IID client distributions, real schools may differ more strongly in curriculum, grading policy, resource availability, student demographics, institutional policy, and data quality. Moreover, the at-risk labels are dataset-specific and partly depend on performance-related variables. As shown in the reduced-feature sensitivity analysis, removing direct outcome-proxy variables leads to lower but still meaningful predictive performance. This indicates that the full-feature results should be interpreted as an upper-bound setting, whereas stricter early-warning deployments require validation using only pre-outcome behavioral, demographic, and engagement indicators. Another limitation concerns privacy, governance, and security evaluation. The membership-inference experiments provide empirical evidence that client-side DP-SGD reduces membership leakage, but they do not exhaust all possible attacks against federated systems. Recent educational FL work emphasizes that DP-based FL should be evaluated not only as a technical mechanism but also as part of a broader compliance and governance framework, including data minimization, consent, institutional accountability, and auditability48. In addition, recent reviews of privacy preservation in FL highlight remaining challenges such as heterogeneous clients, privacy–utility–robustness trade-offs, adaptive adversaries, secure aggregation, and stronger attack-based evaluation protocols49. Therefore, future work should evaluate the proposed framework on naturally partitioned multi-institutional datasets, include additional privacy and robustness attacks, assess fairness and calibration across demographic groups, and explore adaptive privacy mechanisms that tune the noise level according to client size, risk profile, and institutional policy. Conclusion The paper introduced FL-AtRisk-DP-PBT: a federated, privacy-preserving, and verifiable academic at-risk prediction framework across multiple schools. We trained a single global logistic-regression model in three regimes: centralized learning, standard FL, and FL with client-side DP and PBT audit logging. The use of logistic regression was intentional, as it provides coefficient-based interpretability, supports auditability, and reduces computational and communication overhead in federated deployment. We report results on a primary dataset of 14,003 students partitioned into 10 simulated schools, as well as on two external benchmarks. On the primary cohort, the centralized model achieves 99.14% test accuracy, F1 = 0.9915, and AUC = 0.9998; the corresponding federated and DP-enhanced variants yield 98.61%/0.9863/0.9993 and 98.00%/0.9802/0.9992, respectively. These results suggest that, under the current experimental setting, decentralization and privacy-preserving training introduce only a limited reduction in predictive performance relative to the centralized reference model. Cross-dataset experiments on xAPI-Edu-Data and Students Performance in Exams showed similar trends across datasets with different feature spaces and label definitions. Under FL-AtRisk-DP-PBT, the xAPI model achieves approximately 93% accuracy, F1 ≈ 0.92, and AUC ≈ 0.97; the Exams model achieves approximately 94% accuracy, F1 ≈ 0.93, and AUC ≈ 0.98. These findings indicate that the proposed pipeline can maintain competitive performance while keeping student records local and adding privacy-preserving and auditability mechanisms. Because the global model is linear, feature weights remain easy to audit. The feature-importance analysis showed broadly similar coefficient patterns across centralized, FedAvg, and FL + DP+PBT models, with influential variables mainly related to academic performance, engagement, attendance, and study-related behaviors. However, in datasets where the at-risk label is partly defined using performance-related variables, features such as exam scores may act as strong outcome proxies. Therefore, the high predictive performance should be interpreted within the scope of the label definitions and feature sets used in this study. Our ablation of the PBT mechanism shows that cryptographic commitments can be added on top of federated DP-SGD with only a small additional performance reduction and modest computational overhead. FL-AtRisk-DP-PBT provides a transparent and computationally lightweight framework for privacy-preserving student at-risk prediction, but further validation is needed before real-world deployment. Future work should evaluate the framework on naturally partitioned multi-institutional datasets, stricter early-warning settings without direct outcome-related features, and additional analyses of privacy leakage, fairness, calibration, and robustness. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1