Securing End-to-End Reinforcement Learning-Driven Autonomous Driving: A Control Command Utility-Based Intrusion Response System
Qisheng Zhang, Han Jun Yoon, Terrence J. Moore, Seunghyun Yoon, Dong Seong Kim, Hyuk Lim, Frederica F. Nelson, Jin-Hee Cho · IEEE Transactions on Intelligent Vehicles · 2025
End-to-end autonomous driving with deep reinforcement learning (DRL) encounters significant challenges in security and safety, which are critical for the automotive industry's adoption of autonomous technologies. This paper introduces a novel intrusion response system (IRS), called the control command utility-based IRS (CCU), specifically designed for DRL-based autonomous systems. TheCCUprovides a lightweight yet powerful defense against false data injection attacks on the in-vehicle CAN (control area network) bus, enhancing both security and driving performance by making intelligent, context-aware decisions based on control command utilities derived from DRL outputs. We rigorously evaluatedCCUagainst other state-of-the-art IRSs based on DRL autonomous driving models, Rails and Roach. Equipped with an additional confidence score-based filter,CCUeffectively minimizes false alarms, demonstrating superior performance in improving critical driving metrics such as driving score, route completion, and infraction penalties, all while lowering defense costs. Furthermore,CCUexhibits robust resilience in hostile environments with varying attack probabilities, underscoring its reliability in complex scenarios. This contribution represents a significant advancement in autonomous driving, addressing essential security and safety challenges and accelerating the path toward safer, more reliable autonomous vehicle deployment.