Adaptive Federated Learning Framework for Fair and Privacy-Preserving Educational Analytics

The rapid digitalization of educational environments has led to the large-scale generation of student-related data from online learning platforms, intelligent tutoring systems, and virtual classrooms. While such data enable advanced predictive analytics and personalized learning, they also raise critical concerns related to data privacy, fairness, and ethical use. To address these challenges, this paper proposes an Adaptive Federated Learning (AFL) framework for fair and privacy-preserving educational analytics. Unlike conventional centralized learning approaches, the proposed framework allows multiple educational institutions to collaboratively train predictive models without sharing raw student data. Moreover, AFL incorporates a fairness-aware adaptive aggregation mechanism that dynamically adjusts client contributions based on both local predictive performance and fairness indicators. This strategy improves model robustness and reduces demographic bias under heterogeneous and potentially adversarial data distributions. The effectiveness of the proposed framework is evaluated on three publicly available educational datasets—Student Performance, Predict Students’ Dropout and Academic Success, and the Open University Learning Analytics Dataset (OULAD)—under both normal and label-flipping attack scenarios. Experimental results demonstrate that AFL achieves performance comparable to centralized models while consistently improving fairness and resilience against adversarial behavior. These findings highlight the potential of AFL as a trustworthy and ethically aligned solution for decentralized educational data analytics.