Optimized Deep Learning-based Intrusion Detection System for Connected Vehicles
Obinna Agbo, Mohamed Hefeida, Amr S. El-Wakeel · 2025
Future connected vehicles will enable multiple applications on the Internet of Vehicles while maintaining safe transportation. However, securing the vehicles is still a challenging concern for potential future deployment, especially considering intrusion incidents. In this paper, we propose a deep learning-driven intrusion detection system (IDS) that leverages robust and diverse datasets on vehicle intrusions, such as ROAD from Oakridge National Laboratory (ORNL). Unlike existing approaches that rely on datasets that lack sophisticated, stealthy, and complex attack scenarios, our model is evaluated on more comprehensive and challenging threats. This ensures a rigorous assessment of its effectiveness, addressing the limitations of IDS models that struggle to detect advanced intrusions in real time, which potentially poses severe safety risks to vehicles. Using the ROAD dataset, our model achieved 97.85 % accuracy with an MCC of 0.6973. To handle class imbalance in the ROAD dataset, we utilized a class weighting strategy that enabled the precise detection of rare attack categories while maintaining an average low false positive rate (FPR) of 0.0036. Random search hyperparameter tuning further optimized performance, yielding an efficient prediction time of 13.26 seconds for 307,838 ROAD dataset test samples and 71.50 seconds for HCRL test samples. These results highlight the potential of our model as a lightweight IDS optimized for resource-constrained environments.