C-Sense: A Wi-Fi Channel State Information and LSTM-Based Deep Learning Framework for Privacy-Preserving Suspicious Human Activity Detection

Abstract

Conventional surveillance infrastructure relies almost exclusively on camera-based monitoring, which raises privacy concerns, requires adequate illumination and unobstructed line of sight, and depends on continuous human supervision. This paper presents C-Sense, a device-free, privacy-preserving framework for detecting suspicious human activity using Wi-Fi Channel State Information (CSI) and a Long Short-Term Memory (LSTM) deep learning model. CSI measurements were acquired using low-cost ESP32 microcontrollers and processed through a pipeline comprising sliding-window segmentation, thirteen-dimensional statistical, signal-domain, frequency-domain, and Received Signal Strength Indicator (RSSI) feature extraction, and sequence-based dataset generation. A two-layer LSTM network implemented in PyTorch was trained on 4,903 labelled samples spanning three activity categories: Walking, Empty, and Suspicious. On a held-out test set of 970 samples, the proposed model achieved a test accuracy of 98.87% and a weighted F1-score of 98.86% (macro F1-score of 98.48%), with per-class F1-scores of 99.88% (Walking), 98.61% (Empty), and 96.95% (Suspicious), converging to a training accuracy of 98.87% with a final loss of approximately 0.03-0.04. These results outperform conventional Support Vector Machine, Random Forest, and CNN-based baselines reported in the literature. Confusion matrix analysis confirmed strong discriminative capability across all three classes, with negligible misclassification, most notably for the minority Suspicious class despite its comparatively smaller sample count. The results demonstrate that Wi-Fi CSI, combined with temporal deep learning models, constitutes a viable, low-cost, and privacy-preserving alternative to camera-based suspicious activity recognition, particularly suited to indoor security, smart-home, and assisted-living applications.