Evaluating the Stochastic Robustness of Ensemble Learning Architecture for Multiclass Intrusion Detection in CAN-Based IoV Systems

Tamara Zhukabayeva, Lazzat Zholshiyeva, Nurdaulet Karabayev, Yerik Mardenov, Dilaram Baumuratova

The paper investigates the problem of multi-class intrusion detection in CAN-oriented vehicle networks for intelligent transport systems (ITS). The growing interconnectedness of such networks makes them vulnerable to cyberattacks, which requires the development of efficient and productive detection systems. The CIC IoV 2024 dataset is used, which contains CAN-bus traffic in both normal operation mode and various attack scenarios. The classification task includes the following classes: normal traffic, DoS attacks, as well as parameter change attacks for gas (GAS), engine speed (RPM), car speed (SPEED) and steering wheel control (STEERING_WHEEL). Data preprocessing involves scaling traits and applying the Principal Component (PCA) method to reduce dimension and eliminate correlations between traits. Detection system implemented using Decision Tree, Random Forest and AdaBoost algorithms. Performance assessment is based on the macro-averaged metrics Accuracy, Precision, Recall and F1-score. To assess the model’s resilience under real cyber-physical attack conditions, a stochastic sensitivity test is offered in which stochastic perturbations are added to incoming CAN traffic. This allows simulating real-world disturbances, as opposed to optimizing Stochastic Sensitivity Tests. The comparison of clean and noisy data is done using error matrices and degradation analysis of quality indicators. Experimental results show that emsemble methods are effective in increasing the stability of intrusion detection systems in CAN-routed networks.