Multi-Class Intrusion Detection in Internet of Vehicles: Optimizing Machine Learning Models on Imbalanced Data

Ágata Palma, Mário Antunes, Jorge Bernardino, Ana Alves

The Internet of Vehicles (IoV) presents complex cybersecurity challenges, particularly against Denial-of-Service (DoS) and spoofing attacks targeting the Controller Area Network (CAN) bus. This study leverages the CICIoV2024 dataset, comprising six distinct classes of benign traffic and various types of attacks, to evaluate advanced machine learning techniques for instrusion detection systems (IDS). The models XGBoost, Random Forest, AdaBoost, Extra Trees, Logistic Regression, and Deep Neural Network were tested under realistic, imbalanced data conditions, ensuring that the evaluation reflects real-world scenarios where benign traffic dominates. Using hyperparameter optimization with Optuna, we achieved significant improvements in detection accuracy and robustness. Ensemble methods such as XGBoost and Random Forest consistently demonstrated superior performance, achieving perfect accuracy and macro-average F1-scores, even when detecting minority attack classes, in contrast to previous results for the CICIoV2024 dataset. The integration of optimized hyperparameter tuning and a broader methodological scope culminated in an IDS framework capable of addressing diverse attack scenarios with exceptional precision.