A Risk-Aware Deep Reinforcement Learning Framework for AI-Driven Intrusion Detection and Adaptive Response in Autonomous Vehicles

Volume 18, Number 2

A Risk-Aware Deep Reinforcement Learning Framework for AI-Driven Intrusion Detection and Adaptive Response in Autonomous Vehicles

Authors

Muhammad Faisal Shafiq ¹, Muhammad Irshad ² and Muhammad Naveed Sajjad ³, ¹ Lucid Motors, Saudi Arabia, ² RMG Company, Saudi Arabia, ³ Yanal Finance Company, Saudi Arabia

Abstract

Autonomous vehicles expose in-vehicle networks to sophisticated intrusion threats. Although deep learning-based intrusion detection systems achieve high accuracy, most approaches remain detectioncentric and do not address adaptive mitigation under real-time constraints. This paper proposes a multilayer AI-driven framework integrating temporal deep learning, contextual risk modeling, and deep reinforcement learning-based mitigation. A hybrid CNN–BiLSTM model captures spatial payload characteristics and bidirectional temporal dependencies in CAN sequences. Detection output feeds a riskaware formulation fusing attack probability with ECU criticality, vehicle speed, and safetyindicators. A Deep Q-Network learns mitigation policies minimizing residual system damages. Evaluation on the HCRL car-hacking dataset demonstrates strong detection performance. The adaptive policy reduces average system damage cost by 54.63% compared to static monitoring. End-to-end latency of 0.0205 ms per sample satisfies real-time constraints. By integrating detection, severity modeling, and adaptive response, the proposed architecture advances intrusion detection toward intelligent cyber-physical mitigation for resilient autonomous vehicles.

Keywords

Autonomous vehicles, Controller Area Network (CAN), Intrusion Detection System (IDS), Deep Learning, CNN–BiLSTM, Risk-Aware Modeling, Deep Reinforcement Learning, Adaptive Mitigation, Cyber-Physical Security, Real-Time Systems