WiSafe: a real-time system for intrusion detection based on wifi signals

Pioneer research works for WiFi-based sensing usually depend on an extremely high sampling rate of over 1000 Hz to collect an abundant number of Channel State Information (CSI) measurements, which depict the environmental changes and human motions. However, these methodologies can be hardly deployed on embedded platforms and cannot be directly commercialized. In this paper, we propose WiSafe, a real-time intrusion detection system that works with a relatively low sampling rate of 15--20 Hz on commercial embedded devices. Both primitive physical features and high-level CSI-based features are adopted as the input of the classification methods. We exploit MLP, TextCNN and Bi-LSTM together for majority voting to improve the detection accuracy and robustness. Experiment results demonstrate that WiSafe achieves an accuracy of 97.8% (AUC=0.9931) that is comparable to those of previous works even with a low sampling rate, and can run in real time, which make it permissible for practical use.