Adversarial Attacks on Multi-Agent Reinforcement Learning Systems

As multi-agent reinforcement learning (MARL) is increasingly applied in real world applications, ensuring the robustness against adversarial threats becomes essential. This thesis investigates the vulnerability of cooperative MARL systems to adversarial attacks. This was achieved by implementing two types of attacks, random agent behavior and observation disruptions, on agents trained with the QMIX algorithm, within the PettingZoo Pursuit environment. The study evaluates the impact of these attacks on agent performance and coordination. Additionally, it explores the potential of adversarial training as a possible defense to the attacks. The results indicate that both attacks significantly degrade performance when applied at a system trained under standard conditions. However, agents exposed to attacks during training demonstrated improved performance during evaluation with attacks, while they performed worse under standard conditions. This suggests that adversarial training improves performance under attacks, but that there is a trade-off between performance and robustness.