CAT: Cross-Adversarial Training for WiFi-Based Human Activity Recognition

When attacked by mainstream adversarial attack methods, the classical WiFi-based human activity recognition (HAR) system shows poor robustness. Nevertheless, popular standard adversarial training (SAT) conducts adversarial attacks on every data batch by selecting partial perturbation values, resulting in unfavorable model performance and robustness. To mitigate the issues, we propose cross-adversarial training (CAT), an adversarial training technique that cross-attacks the source domain data in each batch. We develop a filter to establish the perturbation threshold for the parameter (i.e., ϵ) of adversarial examples that are distinguishable to naked eyes. Further, we design cross-zero, which crosses the zero point of ϵ and randomly selects the value of ϵ from positive and negative intervals of the threshold range to generate adversarial examples. Our experiments demonstrate that when utilizing the fast gradient sign method (FGSM) and projected gradient descent (PGD), CAT outperforms SAT by an average of 10.37% and 16.33% in accuracy, while SAT results in a decrease of 6.37% and 12.53% compared to the original model (ORI), respectively. To evaluate robustness, we also use a series of ϵ to attack ORI, SAT, and CAT. Results show that the overall robustness of CAT prevails. We believe CAT has a bright future in WiFi-based HAR applications since it increases the system’s security and efficiency through mitigating prevalent adversarial attacks.