How a backdoor implanted in just one agent of a cooperative multi-agent deep reinforcement learning team can pry the entire system into failure. Covers spatiotemporal behavior patterns as stealthy triggers spread across a sequence of observations instead of instant visual patterns, decoupling the trigger from a controllably delayed attack period, and reward hacking with a unilateral influence filter that amplifies the backdoored agent's influence on teammates while suppressing the reverse influence. Includes target-failure-state guidance mined from historical trajectories, attack results against VDN, QMIX, and MAPPO teams, and resistance to representative backdoor defenses.
BLAST is a backdoor attack against cooperative multi-agent deep reinforcement learning (c-MADRL) systems that compromises an entire team by implanting a backdoor in only a single agent. Instead of fi…