Privacy-Aware Human–AI Collaboration for Cross-Organizational Intelligence Using Federated Learning

Abstract

This study evaluates privacy-preserving adaptive learning analytics under a cross-silo federated learning setting that combines secure aggregation, update-level differential privacy, and lightweight personalization to address non-IID institutional data. Experiments across 8–20 institutional clients and participation rates of 0.30–0.70 show that secure aggregation preserves utility while differential privacy introduces a controlled tradeoff. Global discrimination remains competitive at moderate privacy strength, with AUC decreasing from 0.861 (σ=0.2) to 0.801 (σ=2.0) and macro-F1 decreasing from 0.798 to 0.721 over the same range, while variability increases from 0.006 to 0.011 in AUC standard deviation. Decision reliability degrades with stronger privacy, with ECE increasing from 0.033 (σ=0.2) to 0.074 (σ=2.0) and high-confidence error rising from 0.09 to 0.26, indicating elevated risk for threshold-based interventions. Personalization improves both utility and equity under heterogeneity, increasing mean AUC from 0.832 to 0.857 and macro-F1 from 0.741 to 0.782, while reducing client dispersion (AUC std 0.041→0.033; macro-F1 std 0.056→0.041) and improving calibration (ECE 0.055→0.048). Communication analysis shows that rounds-to-target decrease materially with participation and balanced local computation, reaching 86 rounds at p=0.70 with E=3 compared with 156 rounds at p=0.30 with E=1. Overall, the results demonstrate that privacy-preserving federated analytics can remain intervention-grade when privacy parameters, participation scheduling, and calibration-aware personalization are jointly governed.