Lightweight Privacy-Preserving Human Activity Recognition from CSI Data using CNN-Temporal Attention Network

Abstract

WiFi Channel State Information (CSI) has emerged as a powerful sensing modality for device-free Human Activity Recognition (HAR), enabling fine-grained motion understanding without requiring wearable sensors or cameras. However, any type of HAR -- either CSI signal-based or device-based, inherently encode sensitive behavioral patterns, raising significant privacy concerns. In this work, we propose an end-to-end privacy-preserving CSI-based HAR framework that integrates a Convolutional Neural Network (CNN) with a temporal attention mechanism. We perform extensive evaluations on multiple benchmark datasets consisting of varying distance and height factors, as well as different environmental conditions. Our baseline non-privacy-preserving CNN--Temporal attention model achieves state-of-the-art performance. Additionally, we incorporate differential privacy (DP) into the training pipeline -- enabling rigorous privacy guarantees through controlled noise injection and gradient clipping. We evaluate the proposed framework's privacy--utility trade-off and demonstrate that even a strong privacy protection can maintain excellent recognition accuracy. Our framework can progressively approach the non-privacy-preserving performance for some parameter regime. As such, our experimental results clearly demonstrate that the proposed architecture remains robust under privacy constraints and generalizes effectively across heterogeneous sensing conditions. We argue that our work provides practical insights into deploying secure and privacy-aware WiFi sensing systems for real-world HAR applications.