pFedRadar: Personalized Federated Learning for Radar-Based Human Activity Recognition
Abstract
Radar-based human activity recognition offers a noncontact solution for IoT-enabled smart environments but faces data insufficiency challenges in real-world deployments. Although aggregating radar data from multiple devices and users can mitigate the limitations of single-source data, sharing raw radar data is often impractical due to privacy concerns, particularly when it contains personal or sensitive information. Considering the inherently non-IID nature of radar data from different nodes, this article addresses the dual heterogeneity of label and feature distributions in radar-based human activity recognition by proposing a personalized federated learning (PFL) framework, termed pFedRadar. To mitigate label distribution heterogeneity, we introduce a class-aware structural similarity weighting scheme that enables personalized aggregation based on interclient relationships. To alleviate feature distribution heterogeneity, we design a learnable historical knowledge fusion mechanism that mitigates the “round drop” phenomenon across communication rounds in federated learning (FL)-based training. The effectiveness of the proposed pFedRadar framework is validated by constructing radar-based human activity datasets under two distinct non-IID settings. Comprehensive experiments against 11 state-of-the-art baselines demonstrate that pFedRadar achieves superior performance in dual non-IID scenarios.