Advancing Intra-Vehicular Communication Security Using Seq2Seq Transformers and Q-Learning for Cyber-Attack Detection

Abstract

The increasing reliance on connected systems and autonomous features in modern vehicles has brought significant advancements in functionality and safety but has also introduced critical cybersecurity challenges, particularly in the Controller Area Network (CAN) bus. This work addresses these challenges by developing a robust framework for real-time intrusion detection using Seq2Seq Transformer architectures and Q-learning. The objective is to enhance detection accuracy and response speed for various cyber-attacks, including Denial of Service (DoS) and message spoofing, ensuring the secure operation of vehicular communication networks. The methodology integrates advanced feature engineering techniques, such as message frequency and temporal dynamics, to capture complex sequential patterns in CAN bus data. The Seq2Seq Transformer model identifies intricate attack patterns, while Q-learning optimizes decision-making for dynamic intrusion scenarios. Using the Car Hacking Dataset, the proposed framework is rigorously evaluated on key metrics, including precision, recall, F1 score, and detection time. Results demonstrate high detection accuracy with an F1 score exceeding 95%, reduced false alarms, and faster response times compared to traditional methods. By combining adaptive learning capabilities and powerful sequence modeling, this work establishes a scalable and effective solution to automotive cybersecurity challenges, paving the way for secure and intelligent connected vehicle ecosystems.