Xinwen Hou

About

Xinwen Hou is a researcher specializing in the security and robustness of reinforcement learning (RL), with a particular focus on offline RL systems. His major contributions center on exposing and mitigating vulnerabilities in RL pipelines, most notably through his work on backdoor attacks in offline RL datasets. In his highly cited paper "Baffle: Hiding Backdoors in Offline Reinforcement Learning Datasets" (2024, 5 citations) and its earlier version (2022, 3 citations), Hou demonstrated how malicious data providers can stealthily embed backdoors into pre-collected offline RL datasets, enabling adversaries to trigger harmful agent behaviors during deployment. This research is critical for the safety and trustworthiness of offline RL, which is increasingly used in real-world applications where direct environment interaction is costly or dangerous. Hou’s work has garnered attention for revealing a previously underexplored attack surface, earning him recognition as a key voice in RL security. His findings serve as a wake-up call for the RL community, urging the development of robust defenses and dataset verification methods to safeguard against such hidden threats.

Research Focus

Key Achievements

Top Papers

1. 1
   Baffle: Hiding Backdoors in Offline Reinforcement Learning Datasets

   5 citations · 2024
2. 2
   BAFFLE: Hiding Backdoors in Offline Reinforcement Learning Datasets

   3 citations · 2022

Key Collaborators

Contact & Links