Securing Multi-Agent Systems Against Hidden Byzantine Attacks.

Cooperative multi-agent systems are vulnerable to sophisticated "Byzantine attacks," where a subset of agents, unknown to the system, can covertly modify their contributions to a joint action after observing the team's plan. This stealthy manipulation poses a significant challenge to system security and performance, as the central learner cannot directly observe these overwrites or the final executed action. This study delves into the theoretical foundations of securing such systems. It first characterizes how the attacker's information—specifically, whether they observe the planned action—shapes the problem's geometry, leading to different robust Markov Decision Process (MDP) models. The research then establishes the fundamental information-theoretic limits of security learning, demonstrating that achieving optimal security regret is inherently tied to a "cumulative response gap" that cannot be avoided. To address this, the paper introduces a novel "stage-tied robust estimation-to-decisions learner." This algorithm is designed to optimize team performance against the worst-case overwrites, aiming for the optimal security value. The researchers provide a theoretical proof for the algorithm's regret bound, offering a comprehensive framework for building reliable multi-agent systems in the presence of hidden adversarial behavior.