Pilot Backdoor Attack Against Deep Reinforcement Learning Empowered Intelligent Reflection Surface for Smart Radio

• Published in: IEEE transactions on wireless communications Vol. 24; no. 6; pp. 4891 - 4903
• Main Authors: Huang, Yunsong, Wang, Hui-Ming, Wang, Zhaowei, Liu, Weicheng
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.06.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Array signal processing; Artificial intelligence; backdoor attack; Beamforming; Communication system security; Data transmission; Deep learning; Deep reinforcement learning; intelligent reflection surface; Optimization; Partial transmit sequences; pilot contamination attack; Principal component analysis; Reflection; Training; Transmission efficiency; Vectors; Wireless communication
• ISSN: 1536-1276; 1558-2248
• DOI: 10.1109/TWC.2025.3545066

Intelligent reflection surface (IRS) has been used to assist communication by reflection and beamforming where a direct path is not available. Thus, IRS adjusts the wireless channel to enhance the data transmission efficiency with low power consumption. Recently, deep reinforcement learning (DRL) has been exploited in IRS coefficients optimization. IRS can be controlled by DRL to adapt their phase shift to the propagation environment and an expected reflection pattern can be obtained. However, due to the openness of wireless channel and the unexplainability of DRL, it is vulnerable to adversary attacks lauched through wireless channels. In this paper, we investigate a pilot contamination based backdoor attack against DRL based IRS beamforming, where an IRS controlled by an adversary attacker is used to contaminate the channel state information (CSI) during the training phase. The backdoor attack is covert because the adversary need not to know the legitimate pilot sequence and DRL performs well when the adversary IRS keeps inactive. We show that the backdoor attack can reduce the data rate significantly with the adversary IRS. At last, we propose a retraining method agaist the attack to recover the data rate.