DriftErase-HAR: Ledger-Anchored Federated Unlearning for Multimodal Human Activity Recognition under Asynchronous Sensor Drift

The paper proposes DriftErase-HAR, a novel federated unlearning framework designed to address the intertwined challenges of sensor drift and data deletion in multimodal human activity recognition. The framework introduces a compact, drift-adaptive backbone using mask-aware multimodal encoding, gated low-rank adapters, and coverage-aware attention-based fusion. It further presents a three-stage unlearning mechanism combining partition-weighted ledger rollback, distribution-matching data scrubbing, and fairness-aware model repair, supported by a hash-chained update ledger for auditability. Experiments on the WISDM dataset demonstrate that DriftErase-HAR substantially reduces membership-inference leakage while improving recognition accuracy and preserving client-level fairness. Moreover, it achieves near-identical parameterization to the retraining oracle with 22.79× fewer optimizer steps, demonstrating an effective trade-off between privacy, utility, fairness, auditability, and computational efficiency.