Backdoor Attacks on Multi-Agent Reinforcement Learning-based Spectrum Management

Effective spectrum management control through multi-agent deep reinforcement learning holds promising potential for advancing wireless communication systems. However, backdoor attacks can compromise the integrity and security of multi-agent deep reinforcement learning models, allowing attackers to manipulate their behaviour and cause significant damage to the system. In this paper, we have defined a four-step process for designing a general backdoor in spectrum management based on multi-agent deep reinforcement learning, which involves searching for the most observed channels, determining the backdoor power limit, selecting feasible poisoned channels, and setting up induced rewards. Experimental results demonstrate the effectiveness of this attack, which allows the system to perform spectrum management without triggering the backdoor. However, when the backdoor is triggered, it results in severe communication interruptions. Overall, this paper contributes to the field of secure and reliable spectrum management by providing insights into the impact of backdoor attacks on deep learning-based systems.