Backdoor Detection in Reinforcement Learning Agents for Electric Vehicle Charging Control
Ajay Raghavan
Affiliation: Eastlake High School
IJSCAR Vol. 3, Issue 2 (2026) · pp. 40–46
DOI: 10.67149/yhjs2024.5/t8m6z3qp
Abstract
As electric vehicles become central to modern transportation power grids increasingly rely on automated reinforcement learning controllers. This study investigates whether backdoored RL agents controlling simulated EV charging systems can be detected using lightweight statistical anomaly detectors and compact neural models. We evaluate detection methods operating solely on state-action trajectories without access to model internals. Across multiple random seeds and held-out evaluation runs neural classifiers achieved strong separation between clean and compromised agents in the evaluated trigger setting while statistical methods exhibited high recall but elevated false alarm rates. Additional robustness experiments with subtle-action probabilistic delayed-effect and stealthy adaptive variants show that performance remains high but slightly weakens under harder attacks mainly through increased false alarms. These results suggest that trajectory-level behavioral monitoring is promising but broader testing under more realistic and adversarially optimized conditions is needed before making general claims about RL backdoor detection.
Keywords: Cybersecurity, Backdoor Detection, Reinforcement Learning, Electric Vehicles