Privacy-Aware Hybrid Federated Learning Framework with Clustered and Quantum Approaches for Wearable Based Stress Detection

Abstract

This work presents a hybrid federated learning framework for stress detection using physiological signals collected from wearable sensor devices. The system utilizes the WESAD dataset, which includes multimodal physiological signals such as electrocardiogram (ECG), electrodermal activity (EDA), respiration, and body temperature. A binary classification task is performed to distinguish stress and non- stress conditions after signal preprocessing, normalization, and feature preparation. Feature selection is carried out using a Random Forest classifier to identify the most relevant physiological attributes and reduce dimensionality. A centralized Multilayer Perceptron (MLP) model is first developed as a baseline, followed by a Federated Learning (FL) framework in which multiple client models are trained locally without sharing raw data. To address heterogeneity in client data distributions, Clustered Federated Learning (CFL) is introduced by grouping clients based on model similarity before aggregation. In addition, a Quantum Federated Learning (QFL) approach is incorporated to further enhance the learning framework for stress classification in privacy-preserving settings. Experimental evaluation is conducted under centralized, federated, clustered, and quantum federated settings, and performance is measured using accuracy, precision, recall, and F1-score. The framework demonstrates effective stress detection performance while preserving privacy by ensuring that sensitive physiological data remains decentralized.