FedPure: Data poisoning attack detection and purification for federated skeleton-based action recognition

Skeleton-based action recognition (SAR) often requires centralized skeleton data, raising serious privacy concerns in deployment scenarios such as healthcare or surveillance. Federated learning (FL) allows SAR models to be trained without sharing raw data and has therefore become an attractive approach for privacy-sensitive, distributed applications such as camera-enabled devices, human–robot interaction, and security monitoring. However, FL-based SAR remains vulnerable to data poisoning attacks. We propose a data poisoning attack detection and purification method for federated SAR, called FedPure. FedPure introduces a fused transform prototype representation, which combines global perspective transforms with subregion transforms to capture spatiotemporal cues. This design enables precise inter-client correlation analysis for malicious client detection. Moreover, a detector comprising inter-client spatiotemporal matching is designed to analyze the correlation between pseudo skeleton data. Furthermore, FedPure improves model robustness by purifying the malicious clients using a disentangled feature-based purifier to maintain data diversity. The experimental results on diverse adversarial attacks, including FGSM, PGD, C&W, Bone Length Attack, and Hard No Box Attack, confirm that FedPure outperforms existing models in SAR accuracy. By providing an integrated detection-and-purification pipeline tailored to federated SAR, FedPure narrows a key gap in privacy-preserving training, enabling safer application of FL-based action recognition. Our code is publicly available at https://github.com/alsgur0720/fedpure.