Angel or Demon: Investigating the Impact of Plasticity Interventions on Backdoor Threats in Deep Reinforcement Learning

The authors conduct an empirical study of 14,664 cases and find that most plasticity interventions in modern deep reinforcement learning agents can mitigate backdoor threats, and only SAM exacerbates backdoor threats. Pathological analysis attributes the exacerbation to backdoor gradient amplification and the mitigation to activation pathway disruption and representation space compression. The authors thereby propose the SCC conceptual framework for robust backdoor injection and identify loss landscape sharpness anomalies as a key indicator for DRL backdoor detection.