<ul><li>Offline Reinforcement Learning (RL) is vulnerable to poisoning attacks due to its reliance on externally sourced datasets.</li><li>Certified defenses have been extended to provide larger guarantees against adversarial manipulation in RL.</li><li>The approach leverages properties of Differential Privacy to ensure robustness in both continuous and discrete spaces as well as stochastic and deterministic environments.</li><li>Empirical evaluations show that the approach significantly improves performance under poisoning attacks compared to prior work, enhancing safety and reliability in offline RL.</li></ul>

Multi-level Certified Defense Against Poisoning Attacks in Offline Reinforcement Learning

Discover more