Federated learning (FL) offers a promising solution for human activity recognition (HAR) in healthcare by enabling model training on decentralized data, thereby preserving privacy in compliance with regulations such as GDPR and HIPAA. This study investigates the privacy vs performance trade-offs of FL with centralized machine learning (CML) using the UCI HAR dataset. We focus on three aggregation methods: federated averaging (FedAvg), federated proximal (FedProx), and Krum, under both independent and identically distributed (IID) and non-IID data settings. We evaluate their robustness to poisoning attacks and the impact of local differential privacy (LDP).

Our results show that FL outperforms CML in HAR tasks. In non-IID settings, FedAvg achieves up to 97\% accuracy, outperforming FedProx (91\%) and Krum (88\%). Interestingly, non-IID data yields better performance across all methods. While Krum demonstrates strong resilience against poisoning attacks in the absence of LDP, FedProx maintains greater stability when LDP is applied. However, higher privacy levels reduce accuracy to 58–65\%. These findings position FedProx as a balanced option for privacy-preserving healthcare HAR, emphasizing the importance of carefully tuning privacy mechanisms to maintain optimal performance.