Privacy-Preserving Learning Analytics for Educational Systems

Abstract

Learning analytics systems that aggregate individual learner interaction data to generate insights -- engagement patterns, dropout risk scores, knowledge state estimates, and performance predictions -- create privacy risks that conventional anonymisation techniques cannot adequately address: re-identification from quasi-identifiers in interaction logs, inference attacks that reconstruct sensitive attributes from behavioural data, and membership inference attacks that determine whether an individual's data was used in a model. Privacy-preserving learning analytics (PPLA) applies cryptographic and statistical techniques -- differential privacy, federated learning, homomorphic encryption, and secure multi-party computation -- to enable educationally valuable analytics while providing formal privacy guarantees that conventional de-identification cannot match. This paper proposes the Privacy-Preserving Learning Analytics Framework (PPLAF), a systematic evaluation of five PPLA techniques across four analytics tasks -- dropout prediction, engagement modelling, knowledge tracing, and competency analytics -- at three institutional scales, involving 84,000 learner records from six partner institutions. PPLAF introduces the Privacy-Utility Balance Score (PUBS) integrating formal privacy guarantee strength, analytics utility preservation, and computational cost. Key results: federated learning with differential privacy achieves the highest PUBS (0.912), preserving 94.8% of centralised analytics utility at (ε=2, δ=10-5) privacy guarantee; local differential privacy achieves the strongest individual privacy guarantee at 84.2% utility preservation; homomorphic encryption enables exact computation on encrypted data at 18.4x computational overhead. The framework provides PPLA technique selection guidance and deployment specifications for educational data stewards.