FedHAR-XAI: A Privacy-Preserving Federated CNN–LSTM Framework for Explainable Human Activity Recognition from Sensor Data

Abstract

Human Activity Recognition (HAR) using wearable sensor data has gained significant attention due to its applications in healthcare, smart environments, and human-computer interaction. However, traditional centralized learning approaches suffer from privacy risks and limited scalability. To address these challenges, this paper proposes a privacy-preserving and distributed HAR framework integrating a multi-branch Convolutional Neural Network (CNN) with Bidirectional Long Short-Term Memory (BiLSTM) and Federated Learning (FL). Furthermore, Differential Privacy is incorporated using the DP-SGD mechanism to ensure data confidentiality during model training. The proposed model effectively captures spatial and temporal dependencies from sensor signals while maintaining privacy. Experimental evaluation on benchmark HAR datasets demonstrates that the proposed approach achieves superior performance compared to baseline models, attaining an accuracy of up to 96.7% while significantly reducing vulnerability to membership inference attacks. The results validate the robustness, scalability, and privacy-preserving capability of the proposed system.