Junfeng Guo Heng Huang

PolicyGuard achieves an impressive average AUROC of 0.859 against adversary-agent attacks, setting a new benchmark for backdoor defense in reinforcement learning.