Intelligent Intrusion Detection System in Internal Communication Systems for Driverless Cars

Abstract

The modern car is a complicated system consisting of Electronic Control Units (ECUs) with engines, detectors, and wired and wireless communication protocols, that communicate through different types of intra-car networks. The cyber-physical design relies on this ECU network that have been evidenced to be susceptible to attacks, by security researchers through physical and remote access to the cars' internal network. The internal network contains several security vulnerabilities that make it possible to launch attacks via buses and propagation over the entire ECU network, for example preventing the engine to work or cutting the brakes by injection fabricated messages. therefore, anomaly detection technology, which represents the security protection, can efficiently reduce security threats. This work proposes anomaly Intrusion Detection System (IDS) using the Artificial Neural Network (ANN) to monitor the state of the car by packets collected from internal buses and to achieve security of the internal network through training the CAN packet to devise the fundamental statistical feature of normal and attacking packets and in defense, extracted the related attribute to classify the attack. Generating new features and using the K-means clustering algorithm to differentiate similar samples and addressing them as a single class to improve the accuracy of the model. Features are evaluated to measure its discrimination ability between classes and to select the best existing features. The experimental study has performed using two sets of data, Simulated data Open Car Test-Bed and Network Experiments (OCTANE) and Real dataset to evaluate our detection system. The experimental results on (OCTANE) demonstrate that the IDS has achieved good performance with a false-positive rate of 1.7%, a false-negative rate of 24.6%, and average accuracy of 92.1%. Experimental evaluation on a real dataset shows that the proposed system has a low false-negative rate of 0.1% and an error rate of 0.006 with an average accuracy of 87.63%