Intrusion Detection System for Autonomous Vehicles Using Non-Tree Based Machine Learning Algorithms

Cynthia Anthony, Walid Elgenaidi, Muzaffar Rao

This research work highlights significant achievements in the domain of intrusion detection systems (IDSs) for autonomous vehicles, which are crucial in enhancing their safety, reliability, and cybersecurity. This study introduces an approach that leverages non-tree-based machine learning algorithms, such as K-nearest neighbors and ensemble learning, to develop an IDS tailored for autonomous vehicles. These algorithms were employed because of their ability to process complex and large datasets with less likeliness for overfitting, their scalability, and their ability to adapt to changing conditions in real time. These algorithms effectively handle imbalanced data, enhancing the detection accuracy of both normal and intrusive instances. The IDS’s performance was validated through the utilization of three real-world datasets, CAN intrusion, CICIDS2017, and NSL-KDD, where the proposed non-tree-based IDS (NTB-MTH-IDS) was measured with the standard measurement metrics: accuracy, precision, F1-score, and recall, including specificity and sensitivity. Notably, the results indicate that K-nearest neighbors and stacking, as part of NTB-MTH-IDS, has an accuracy of 99.00%, 98.57%, and 97.57%, and F1-scores of 99.00%, 98.79%, and 97.54% in the CICIDS2017, NSL-KDD, and CAN datasets, respectively. The results of this research can lead to establishing a robust intrusion detection framework, thereby ensuring the safety and reliability of autonomous vehicles. Through this achievement, road users, passengers, and pedestrians are safeguarded against the consequences of potential cyber threats.