Adversarial Attack Mitigation in Formation Control of Multi-Agent Systems

Approximately 6.1 million car collisions occur annually, with 94% caused by human error. Autonomous Vehicles (AVs) in Multi-Agent Systems (MAS) aim to reduce such incidents. However, MAS, essential for AV coordination, are vulnerable to adversarial attacks due to their decentralized nature. A single compromised vehicle can trigger cascading deviations, leading to system failures. This research proposes a novel framework to enhance MAS resilience against adversarial attacks, ensuring stability in real-time operations like search-and-rescue and disaster relief. The study introduces a dynamic recalibration method that adjusts based on attack detection. When an attack is flagged, the compromised agent’s influence is reduced, and coupling dynamics adjust for unaffected agents. Unlike traditional methods, which isolate detected agents, this approach allows partial information flow to maintain system stability. The framework was validated through simulations of four AV agents in a square formation: one unaltered, one under attack without mitigation, and one with attack prevention. Residual-based detection flagged anomalies, and the mitigation strategy reduced cascading effects. Without prevention, the system suffered a 49% trajectory deviation, while the proposed framework limited deviation to 8%, ensuring stability under attack.