FairErase-FCL: Fairness-Aware Federated Continual Unlearning under Non-IID Educational Data

Abstract

Federated unlearning is commonly evaluated by deletion speed, distance to retraining, and average predictive accuracy. These criteria can miss a consequential post-deletion effect: a client withdrawal may alter the distribution of representation across protected groups. Under non-independent and non-identically distributed educational data, the withdrawn institution may contain a disproportionate share of one group. Reversing its update or resetting deletion-sensitive parameters can therefore change group-specific false-negative rates even when average utility remains stable. This paper introduces FairErase-FCL, a fairness-aware sparse repair method for federated continual unlearning. The method identifies coordinates most sensitive to the withdrawn client, builds a group-stratified buffer from data that remain authorized, and optimizes predictive loss together with a smooth equal-opportunity penalty on a restricted repair mask. We evaluate 14 heterogeneous clients, five temporal cohorts, 26 synthetic features, and 15 paired random seeds. The client with the largest protected-group proportion is removed in each run. FairErase-FCL achieves 78.96% ± 1.85% balanced accuracy, 0.42 percentage points below full retraining. Its equal-opportunity gap is 6.65 percentage points, significantly lower than the 8.05 points produced by full retraining (t(14) = -4.55, p < 0.001), at an estimated 24% of retraining computation. Worst-group accuracy does not improve, revealing a genuine conflict between fairness objectives. The study shows that post-deletion release criteria should report average utility, deletion evidence, and several group metrics rather than accepting a single aggregate accuracy.