Fighting Fire with Fire: Adversarial Evaluation of a Reinforcement Learning Network Alert Inspection System

Researchers conducted a red-team evaluation of a reinforcement learning network alert inspection policy proposed with the participation of the Army Research Lab, and found it robust to an attacker's optimal response. However, by exploiting vulnerabilities in the MDP assumptions of the RL model, the authors found an attack policy capable of overwhelming the defender, then used a double oracle method to retrain the defender with that policy and restore its robustness. The authors state that these two methods can generalize to other RL applications in adversarial environments.