Federated Learning Security in E-Learning: A Clustering-Based Approach to Student Behavior Anomaly Detection and Data Poisoning Mitigation
Year: 2026
Type: conference-paper
Abstract: Artificial intelligence (AI) is being increasingly integrated into Learning Management Systems - particularly Moodle, providing unprecedented potentials for adaptive and data-driven education. However, this advancement brings with it an unaddressed vulnerability: When multiple parties aim to collaboratively train predictive models using Federated Learning (FL), they are susceptible to data poisoning attacks, where a single attacker can silently corrupt the shared global model by flipping student outcome labels. What makes this threat especially insidious in educational applications is that the FL defense mechanisms currently available, which are largely focused on gradient inspection and model parameters, are oblivious to the behavioral semantics of the learner interaction logs. Here we fill this gap. We present a behaviour-aware clustering approach that enables discovery of relevant patterns from anonymized Moodle logs, using federated K-means to detect behavioural anomalies and exploiting the label entropy as a lightweight yet principled anomaly indicator. As far as we are aware, this is among the first behavior-aware defense for data poisoning tailored to federated educational data mining. The conceptual validity of the proposed framework is based on the behavioral coherence of student engagement and achievement in the prior work [1], [13]; full experimental evaluation on real institutional data is reserved for an extended journal version.
Institution: Université Ibn-Tofail
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: hybrid
OpenAlex ID: W7204100262