Abstract
Wi-Fi channel state information (CSI) has emerged as a promising alternative to wearable sensors for passive human activity recognition (HAR). However, its sensitivity to environmental layouts and device heterogeneity leads to highly non-IID data across users. Also, combined with the limited number of labeled CSI samples available on each device, existing federated learning (FL) approaches often suffer from representation drift and poor convergence. To address these challenges, this paper introduces a federated few-shot learning framework that incorporates prototypical learning to promote consistent and discriminative CSI embeddings across clients. Each client extracts magnitude–phase statistical features, conducts episodic N-way K-shot training, and updates its local encoder using a prototypical loss. Unlike FedAvg, the server employs an Adam-based aggregation strategy that integrates prototype-induced parameter updates, providing more stable global alignment under heterogeneous data distributions. Experiments on a multi-configuration CSI dataset demonstrate that the proposed federated prototypical learning method significantly improves recognition accuracy, training stability, and cross-environment generalization compared with state-of-the-art FL baselines. These results highlight the potential of few-shot prototypical learning as a scalable and robust solution for CSI-based human activity recognition.
Funding Information
  • IITP (Institute of Information & Communications Technology Planning & Evaluation)-ITRC
  • Korea government (Ministry of Science and ICT (IITP-2026-RS-2024-00436887)
  • Hankuk University of Foreign Studies Research Fund