A heterogeneous federated learning framework for human activity recognition

Arvin Tashakori

Abstract: Federated human activity recognition (FedHAR) leverages federated learning (FL) to collaboratively train HAR models across clients(users) while preserving privacy by keeping data local to each device. However, FedHAR faces two key challenges: model heterogeneity, arising from the diverse computational capabilities of client devices, and data heterogeneity, resulting from behavioral differences across individual clients. In this paper, we propose HFedHAR, a heterogeneous federated learning framework for HAR, to address dual heterogeneity (i.e. model heterogeneity and data heterogeneity). Under model heterogeneity, clients use different model structures, making gradient-based knowledge sharing ineffective in existing FedHAR approaches. To address this, we first design a new ‘bridge’ (i.e., synthetic data generated from clients with high computational power), to link clients. Each client represents their knowledge as prediction logits on the bridge, enabling structure-agnostic knowledge sharing. To further enhance the bridge’s effectiveness, we introduce an information entropy loss. To tackle data heterogeneity, we employ a similarity-based knowledge distillation strategy based on a relation graph constructed among clients, enabling each client to effectively absorb knowledge from others. We evaluate the proposed HFedHAR framework on four HAR datasets, and experimental results demonstrate its effectiveness in addressing the dual-heterogeneity challenges in FedHAR.