Continuous Gait-Based Identity Authentication via Multi-Activity Domain Incremental Learning

Continuous gait-based authentication offers unobtrusive verification on wearables. Over time, such a system is gradually exposed to the full repertoire of user activities and must adapt accordingly. Domain-incremental learning provides the flexibility and framework for building such applications. We therefore introduce AMRIC (Asymmetric Memory Replay and Identity-aware Contrastive learning), a domain-incremental framework for continuous gait authentication designed to mitigate catastrophic forgetting. Architecturally, AMRIC uses a dual encoder to process identity and activity information separately and integrates three mechanisms: (1) to enable differentiated replay strategies, it maintains separate memory buffers for legitimate and illegitimate user identity features; (2) it uses contrastive learning to keep identity embeddings consistent across activities, making the learning process identity-aware; (3) these components are jointly optimized to balance the learning of new patterns with the retention of old ones. Experiments on two inertial sensor datasets, CDUT-AG and USC-HAD, show that AMRIC significantly reduces the Average Equal Error Rate (AE) and Forgetting Measure (FM) compared to strong baselines, achieving AE reductions from 1.6% to 1.17% and 3.04% to 1.52%, and FM reductions from 1.56% to 1.24% and 2.16% to 1.16%. These results demonstrate the framework’s effectiveness for gait-based identity authentication on wearables in multi-activity, domain-incremental settings.