Backdoor Attack Influence Assessment and Analysis on Multi-Agent Training
Yingqi Cai, Pengna Liu, Dongyue Xu, Xuyan Fan, Yalun Wu, Kang Chen, Liqun Chu, Qiong Li, Endong Tong, Wenjia Niu, Jiqiang Liu · 2023
Recent researches have validated the possibility of exploiting backdoor vulnerabilities in deep reinforcement learning (DRL) systems. However, existing attack methods have limitations in designing effective trigger mechanisms. These methods often rely on heuristic or random search approaches to determine trigger parameters, lacking an automatic approach to evaluate the impact of various trigger configurations on the success of backdoor attacks. This paper introduces a novel backdoor attack method based on Bayesian optimization, which automatically assesses the impact of different trigger configurations on attack effectiveness and provides guidance for optimizing triggers to maximize their impact. We validate our method through extensive experiments conducted on two classical cooperative multi-agent reinforcement learning (CMARL) algorithms, VDN and QMIX, within widely-adopted CMARL gaming platform Star-Craft Multi-Agent Challenge (SMAC). The experimental results demonstrate that different trigger parameter configurations have a significant impact on backdoor attacks.