A Deep Learning Perspective on Connected Automated Vehicle (CAV) Cybersecurity and Threat Intelligence
This work tackles cybersecurity for CAVs, which is critical for vehicle safety and infrastructure, but it is incremental as it applies existing deep learning methods to a specific domain dataset.
This book chapter addresses cybersecurity threats in Connected Automated Vehicles (CAV) by proposing a deep CNN-LSTM model for threat detection, which achieved over 99% accuracy, precision, recall, f1-score, and AUC on the CAV-KDD dataset, outperforming other deep learning algorithms like DNN and CNN.
The automation and connectivity of CAV inherit most of the cyber-physical vulnerabilities of incumbent technologies such as evolving network architectures, wireless communications, and AI-based automation. This book chapter entails the cyber-physical vulnerabilities and risks that originated in IT, OT, and the physical domains of the CAV ecosystem, eclectic threat landscapes, and threat intelligence. To deal with the security threats in high-speed, high dimensional, multimodal data and assets from eccentric stakeholders of the CAV ecosystem, this chapter presents and analyzes some of the state of art deep learning-based threat intelligence for attack detection. The frontiers in deep learning, namely Meta-Learning and Federated Learning, along with their challenges have been included in the chapter. We have proposed, trained, and tested the deep CNN-LSTM architecture for CAV threat intelligence; assessed and compared the performance of the proposed model against other deep learning algorithms such as DNN, CNN, LSTM. Our results indicate the superiority of the proposed model although DNN and 1d-CNN also achieved more than 99% of accuracy, precision, recall, f1-score, and AUC on the CAV-KDD dataset. The good performance of deep CNN-LSTM comes with the increased model complexity and cumbersome hyperparameters tuning. Still, there are open challenges on deep learning adoption in the CAV cybersecurity paradigm due to lack of properly developed protocols and policies, poorly defined privileges between stakeholders, costlier training, adversarial threats to the model, and poor generalizability of the model under out of data distributions.