Wireless Universal Adversarial Attack and Defense for Deep Learning-Based Modulation Classification

Adversarial attacks on deep learning based modulation classification have received considerable attention recently. However, existing works mainly focus on the idealized white-box adversarial attacks and ignore the impact of the wireless channel. In this letter, we present a black-box Universal Adversarial Perturbation (UAP) attack scheme considering the wireless channel and propose the corresponding defense method. We first propose a conditional generative adversarial Nets (cGAN) approach to enlarge the training set of the channel state information (CSI) of wireless channel. Then, we introduce a cGAN aided black-box UAP attack scheme to disable the modulation classification capability of the deep neural network over the air. At last, we present a defense method that utilizes UAPs for adversarial training (AT). Simulation results show that the cGAN aided black-box UAP attack can decrease the accuracy of the modulation classifier by 19.3% when the perturbation power reaches the same level as the noise power, while the proposed defense method can improve it by 11.2%.