BARD-MARL detects rogue AI agents with 0.982 AUC in traffic grids

Multi-agent reinforcement learning (MARL) systems that use learned communication are vulnerable to Byzantine agents — agents that become faulty or adversarial after training. A new paper from Almond Kiruthu Murimi introduces BARD-MARL, a post-hoc diagnostic layer that detects such agents without modifying the underlying MARL algorithm. The approach builds on BayesG (used as an attributed communication substrate) and fuses two evidence streams: policy-graph features extracted from state-action trajectories and Bayesian trust statistics from BayesG's latent mask probabilities.

BARD-MARL was evaluated on SUMO traffic signal control grids with four attack types: fixed-action, observation-flip, random-noise, and coordinated. On a 25-agent grid, it reached 0.843 AUC-ROC under a 10% observation-flip attack, while policy-graph-only detection hit 0.917 AUC-ROC under a 10% coordinated attack. On a 100-agent grid, the unified BARD-MARL variant achieved an impressive 0.982 AUC-ROC for both 10% fixed-action and 10% coordinated attacks. The paper emphasizes that resilience claims require attack-specific ablations and clear separation between coordination, detection, and mitigation phases.