Real-time Adversarial Attack to Deep Learning-based Wi-Fi Human Activity Recognition

This study investigates adversarial attacks on deep learning (DL)-enabled Wi-Fi sensing systems using channel state information (CSI) for privacy. This paper presents a technique to disturb the signal used for channel estimation transmitted from the user device when the classifier is located at the router. We employ generative adversarial imitation learning (GAIL), a deep reinforcement learning method to build adversarial attacks using estimated CSI data without explicit reward function feedback. Our approach does not require the adversary to know the structure or the weights of the network. Our proposed method lowers classifier accuracy to 50% using perturbation signals with an amplitude 1.0 dB lower than those used in the attack scheme based on impractical assumptions.