Privacy-Preserving and Secure AI-based Stress Monitoring Federated Learning on Wearable Edge Devices Using Homomorphic Encryption

Abstract

The paper presents a new framework, HE-FedStress, with which it is feasible to monitor stress and preserve privacy in an AI-based approach to stress detection through federated learning and homomorphic encryption on wearable edge devices. The proposed system handles the essential privacy issues that come with decentralized health surveillance because the proposed system allows joint model training without exposing uncoded biometric information. The Edge wearable devices are local physiological signal temporal attention models that inference latency can process five-second windows with a latency of just 47 ms, implemented on photoplethysmography, electrodermal activity and accelerator data. The updates of the model are coded using the Pailier cryptosystem and sent to the central server where they could be aggregated securely without decrypting each individual contribution. The WESAD and SWELL-KW datasets have been empirically assessed to verify that HE-FedStress can give F1-scores of 89.7% and 85.2% respectively and retain centralized model performance with 92 and 94 % of this performance under full cryptographic protection against model-inversion attacks (compared to 34.7 % centralization in the standard federated learning case). The structure uses gradient quantization, which cuts down the communication load to 63KB/ round, and uses adaptive batch size to support heterogeneous device capacity. The computational optimizations (depthwise separable convolutions, selective attention pruning, and eight-bit quantization) allow incessant 24 hour execution of commercial wearables with only a 19% power consumption impact over plaintext federated learning. The design of the modular architecture is to meet GDPR and HIPAA data-localization standards and provide the ability to extend it to other healthcare applications. Therefore, this contribution can provide a practical, scalable stress monitoring solution to secure and personalized performance with a robust guarantee on cryptography, and can also indicate that the requirements of resource-constrained edge environments can be satisfied with a robust guarantee alongside real-time performance.