BehaviorGuard: Online Backdoor Defense for Deep Reinforcement Learning

A new framework called BehaviorGuard proposes an online, trigger-agnostic defense against backdoor attacks in deep reinforcement learning (DRL). Unlike existing methods that rely on reward anomalies and model fine-tuning, BehaviorGuard detects backdoors by monitoring shifts in action distributions, even without triggers. It identifies suspicious behavior in high-quantile regions and distribution tails, then suppresses backdoor actions at runtime. The approach aims to reduce costs and improve robustness against complex trigger patterns. The paper is published on arXiv under ID 2605.05977.