Attacking cooperative multi-agent reinforcement learning by adversarial minority influence.

Area of Science:

• Artificial Intelligence
• Robotics
• Game Theory

Background:

• Cooperative multi-agent reinforcement learning (c-MARL) systems face vulnerabilities under adversarial attacks, impacting worst-case performance.
• Existing observation-based attacks often rely on impractical white-box assumptions and fail to account for complex multi-agent dynamics.

Purpose of the Study:

• To propose a practical and potent black-box adversarial attack method specifically designed for c-MARL systems.
• To address the limitations of current attacks by considering intricate agent interactions and cooperative objectives.

Main Methods:

• Introduced Adversarial Minority Influence (AMI), a black-box attack requiring no knowledge of victim parameters.
• AMI leverages agent-wise relation metrics derived from mutual information to maximize adversarial impact.
• A reinforcement learning agent determines targeted detrimental scenarios through trial-and-error for long-term cooperative harm.

Main Results:

• Demonstrated the first successful adversarial attack against real-world robot swarms.
• Effectively induced worst-case cooperative scenarios in simulated environments like Starcraft II and Multi-agent Mujoco.
• AMI enables a single adversary to mislead multiple agents into collectively detrimental outcomes.

Conclusions:

• Adversarial Minority Influence (AMI) presents a significant advancement in understanding and exploiting c-MARL vulnerabilities.
• The proposed method offers a practical and effective approach for evaluating the robustness of c-MARL systems.
• Findings highlight the need for enhanced security measures in cooperative AI systems operating in critical domains.