CGTNET-Fed: A Privacy-Preserving Federated Learning Framework For Personalized Human Activity Recognition in Health Monitoring

Abstract

Centralized Human Activity Recognition (HAR) systems play a key role in health monitoring but are highly susceptible to privacy attacks because sensitive sensor data are transferred to a central server. In this study, we present a privacy-preserving federated learning system for activity recognition in decentralized environments and personalized health advice by applying CGTNet-Fed (CNN-BiGRU-Transformer federated model). Here, training is performed in a distributed manner over several clients based on Federated Averaging (FedAvg) and proximal regularization (FedProx) without transferring the raw sensor data to any other location. This framework was tested in a realistic non-IID setting with activities simulated using a Dirichlet distribution in PAMAP2 (12 activities) and UCI HAR (six activities). The experiments showed that federated training could be competitive with only a slight degradation compared with centralized training (In UCI HAR, the Macro F1 scores for federated and centralized were 0.927 and 0.918, respectively; In PAMAP2, they were 0.825 and 0.801, respectively). Under non-IID heterogeneous data, FedProx achieved approximately half of the performance improvement compared to the classical SVM (0.843). The personalization module is lightweight and transforms activity predictions into user-specific health recommendations while preserving privacy. The outcomes provide a reproducible benchmark for federated HAR, showing that the preservation of privacy does not entail a significant utility loss and validating the framework for scalable General Data Protection Regulation-compliant health monitoring in Internet of Medical Things (IoMT) contexts.

Centralized Human Activity Recognition (HAR) systems play a key role in health monitoring but are highly susceptible to privacy attacks because sensitive sensor data are transferred to a central server. In this study, we present a privacy-preserving federated learning system for activity recognition in decentralized environments and personalized health advice by applying CGTNet-Fed (CNN-BiGRU-Transformer federated model). Here, training is performed in a distributed manner over several clients based on Federated Averaging (FedAvg) and proximal regularization (FedProx) without transferring the raw sensor data to any other location. This framework was tested in a realistic non-IID setting with activities simulated using a Dirichlet distribution in PAMAP2 (12 activities) and UCI HAR (six activities). The experiments showed that federated training could be competitive with only a slight degradation compared with centralized training (In UCI HAR, the Macro F1 scores for federated and centralized were 0.927 and 0.918, respectively; In PAMAP2, they were 0.825 and 0.801, respectively). Under non-IID heterogeneous data, FedProx achieved approximately half of the performance improvement compared to the classical SVM (0.843). The personalization module is lightweight and transforms activity predictions into user-specific health recommendations while preserving privacy. The outcomes provide a reproducible benchmark for federated HAR, showing that the preservation of privacy does not entail a significant utility loss and validating the framework for scalable General Data Protection Regulation-compliant health monitoring in Internet of Medical Things (IoMT) contexts.