Prototype-Centered Residual Modeling for Federated Wireless Human Activity Recognition

Wireless sensing-based human activity recognition (HAR) is a noncontact and privacy-preserving technology that has attracted extensive interest in academia and industry. However, in practical deployments, heterogeneity across clients stemming from user habits, environmental dynamics, and sensing conditions severely weakens the generalization of isolated local models. In federated wireless sensing, clients share the same activity semantics, while the radio frequency (RF) observations may vary substantially across environments, placements, and channel conditions. In this article, we propose BridgeFi, a federated learning (FL) framework that addresses this challenge by modeling the discrepancy between local features and shared class prototypes from a distributional perspective. Specifically, BridgeFi treats shared class prototypes as semantic anchors and represents local class features as prototype-centered residual distributions, enabling the degree and shape of cross-client feature-prototype discrepancies to be optimized at multiple representation depths. This formulation regularizes local optimization directions, reduces cross-client inconsistency, and remains compatible with different wireless sensing backbones. Experiments on four public wireless sensing datasets show that BridgeFi improves cross-client collaboration and consistently strengthens out-of-domain recognition.