CAREER: Securing Deep Reinforcement Learning

Like many other deep learning techniques, deep reinforcement learning is vulnerable to adversarial attacks. This project intends to robustify deep reinforcement learning by integrating and expanding upon a series of technical approaches used in explainable Al, adversarial training, and formal verification in conjunction with program synthesis. lf successful, the project will significantly advance the field of Al security (for adversarial training and adversarial policy learning) and contribute to the field of machine learning (for explainable Al and verified Al). The expected advancements are three folds: (1) We will expand an explainable-Al-based adversarial policy learning method to scrutinize reinforcement learning agents and unveil their policy flaws. (2) We will design new methods to improve noise-resiliency for broadly adopted interpretable Al techniques. (3) We will develop complementary techniques to remediate the policy weakness of reinforcement learning agents and improve their robustness.