A Quantum-Secured Federated Deep Learning Framework for Multimodal Stress Prediction Using Smartwatch and Smartphone IoT Data

Multimodal sensing from smartwatch and smartphone IoT devices enables continuous monitoring of physiological and behavioral signals for stress prediction. However, existing edge-AI solutions face challenges related to user privacy, secure communication, and efficient on-device computation. This study presents a Quantum-Secured Federated Deep Learning (QS-FDL) framework that integrates lightweight multimodal learning with quantum-resilient communication using the BB84 and E91 protocols. The model employs a hybrid Multi-Input One-Dimensional Convolutional Neural Network (MI-1D-CNN) and Bidirectional Long Short-Term Memory (Bi-LSTM) architecture optimized for resource-constrained wearable devices. Experiments on the WISDM and WESAD datasets demonstrated high accuracy and robustness while maintaining low computational overhead. The proposed framework enhances confidentiality, reduces communication cost through hierarchical aggregation, and provides interpretable stress predictions using Shapley Additive Explanations (SHAP). Experiments on WISDM and WESAD achieved up to 74.6% and 69.8% feature reduction and classification accuracies of 97.1% and 98.4%, respectively, while reducing computational cost by over 39% compared to existing federated approaches. These results show that QS-FDL is suitable for secure real-time stress analytics across heterogeneous IoT environments.