GAN-Enabled Robust Backdoor Attack for UAV Recognition

Abstract

Unmanned aerial vehicle (UAV) recognition is of crucial importance due to the blowout amount of UAVs and their threats on the public safety. Although many UAV recognition methods based on deep learning (DL) have been proposed by utilizing the radio frequency fingerprints and have achieved appreciable results, their vulnerability to adversarial attacks, especially backdoor attacks, has not been studied. In this pa-per, in order to reveal the serious threat for DL-based UAV recognition encountered with backdoor attacks, a novel robust generative adversarial network (GAN)-enabled backdoor attack scheme is proposed. Moreover, the proposed GAN-based trigger generator not only emerges exceptional attack effectiveness, but also performs well in terms of attack stealthiness and migration ability. Simulation results obtained with the real collected UAV recognition dataset demonstrate that our proposed scheme outperforms the benchmark BadNets backdoor attack.