Cyber attacks and detection mechanisms for driverless cars

Abstract

The introduction of Vehicle-to-Everything (V2X) communications is a fundamental requirement for the evolution of today’s Autonomous Driving, but it leads to a new set of vulnerabilities in network infrastructure. It is important to note that cyber-attacks, including the availability ones, such as DoS, represent a significant threat to the safety of Intelligent Transport Systems (ITS). Traditional signature-based Intrusion Detection Systems (IDS) have a disadvantage in security due to their inability to adapt and manage these new and evolving attacks: they can be blind to new or “zero-day” kinds of attacks. This project is to solve this problem by proposing and validating an unsupervised Intrusion Detection System using a Deep Autoencoder architecture. Unlike typical supervised models, where labelled attack data is needed, this system is trained on normal network traffic patterns only. It tracks anomalies by learning to compress and reconstruct legitimate traffic features, marking large reconstruction errors as malicious intrusions. The model was developed in TensorFlow and tested against the KDD Cup 99 benchmark dataset. Experimental results show the high performance of the system with a total Accuracy of 99.49% and a critical Recall of 99.86%, effectively suppressing almost all availability attacks. In addition, the model is consistent with a Matthews Correlation Coefficient (MCC) of 0.9530, confirming its robustness and reliability even for very asymmetric network traffic. This research establishes solid proof-of-concept for the concept that unsupervised deep learning can work as a powerful new mechanism of security architecture for V2X infrastructure without relying on prior knowledge about specific attack signatures.