Federated attention-driven client selection and aggregation for privacy-preserving human activity recognition

Abstract Human Behavioral Context Recognition (HBCR) aims to identify user activities in real-world environments and has critical applications in assistive living, healthcare, and industrial automation. A major challenge in this field is achieving high model performance while ensuring user privacy, which is difficult for traditional centralized systems. Federated Learning (FL) addresses this by enabling decentralized model training directly on user devices. However, its success depends heavily on effective client selection and aggregation strategies. In this paper, we propose FAAM (Federated Averaging with Attention Mechanism), a novel framework that enhances FL by identifying and prioritizing the most relevant clients during each communication round using an attention-based approach. FAAM optimizes client selection and communication efficiency, resulting in faster convergence, reduced communication overhead, and equivalent model accuracy. We evaluate FAAM on four newly collected Human Activity Recognition (HAR) datasets, demonstrating that it outperforms state-of-the-art techniques in convergence speed and accuracy. Our results show that FAAM achieves optimal performance with fewer clients and communication rounds, making it an efficient and privacy-preserving solution for real-time HBCR systems.