Yunpeng Bai

Papers

Total Citations

H-Index

About

Yunpeng Bai is a researcher focused on the security and robustness of reinforcement learning (RL), with a particular emphasis on offline RL paradigms. His major contributions center on exposing vulnerabilities in offline RL systems, where agents learn from pre-collected datasets rather than real-time environment interactions. Bai’s most notable work, "BAFFLE: Hiding Backdoors in Offline Reinforcement Learning Datasets" (2022, 3 citations), and its 2024 follow-up (5 citations), introduce a novel attack vector: backdoor triggers embedded in offline datasets that can covertly manipulate agent behavior. This research highlights critical security risks in data-sharing practices for offline RL, a paradigm valued for its efficiency in avoiding costly environment interactions. By demonstrating how malicious data providers can compromise learned policies, Bai’s work underscores the need for robust dataset verification and defense mechanisms. His findings are particularly relevant for safety-critical applications of RL, such as autonomous driving and healthcare, where undetected backdoors could lead to catastrophic failures. Bai’s contributions advance the understanding of adversarial threats in machine learning, making him a key voice in the emerging field of RL security.