Fedstress: A Privacy-Preserving Federated Learning Framework for Efficient and Accurate Stress Detection Using Wearable Sensors

The proposed privacy-sensitive federated learning method is FedStress, which is aimed at identifying wearable stressors with high accuracy and efficiency. The growing use of wearables in health monitoring poses serious issues regarding data privacy and computational limitations especially in a situation where sensitive physiological data is to be used. The approaches in existence have inherent trade-offs: federated learning opens model parameters to gradient inversion attacks and differential privacy errors the accuracy by injecting noise and homomorphic encryption is prohibitive due to resource-constrained devices. To solve these issues, FedStress combines federated learning with an optimized homomorphic encryption and allows collaborative model training without having access to raw user data. Every device trains locally a lightweight stress detector using a variant of MobileNetV3 using depthwise separable convolutions and a hybrid attention mechanism in order to trade off accuracy and efficiency. The encrypted transmission of model updates is done through partially homomorphic encryption, sensor-aware ciphertext packing which minimises overhead on encrypted model updates, and allows secure aggregation in the encrypted space. An additional defense mechanism is on-device differential privacy with adaptive noise scaling, which prevents inference attacks and hierarchical key management which implements strict access control. Experimental confirmation of the WESAD dataset shows that FedStress can have 87.6% accuracy in detecting stress as centralized methods (89.1) with a much lower level of privacy risk, a gradient inversion attack success rate of 11% and a score of 0.09 on information leakage (0.45 with standard federated learning). This framework is practical, with 23ms inference time and 0.45J of energy used to make a classification on commercial smartwatch systems due to efficient parameter synchronization and model compression. We also provide a viable solution to the current trend of privacy-conscious health monitoring systems, which offers a reasonable compromise between algorithmic performance and practical functionality of wearable applications. The modular structure also enables it to increase the physiological sensing tasks, which makes it a flexible instrument in studying edge-based federated learning.