An Autoencoder-Based Anomaly Detection Framework for Vehicular Networks Using the CICIoV2024 Dataset

In vehicular networks, the detection of malicious traffic presents significant challenges due to the critical nature of automotive systems and the evolving complexity of cyber attacks. This paper presents an unsupervised anomaly detection framework utilizing autoencoder neural networks for identifying malicious Controller Area Network (CAN) bus traffic. The proposed methodology introduces a novel data preprocessing approach where feature standardization is computed exclusively from benign traffic samples, thereby preventing statistical contamination from attack patterns. Leveraging the CICIoV2024 dataset, which comprises real-world CAN bus traffic collected from a 2019 Ford vehicle, we implement an autoencoder architecture trained solely on benign traffic patterns. The model demonstrates robust performance in distinguishing anomalous behaviors, achieving 96% accuracy, 76% recall, and 83% precision across multiple attack vectors, including Denial-of-Service (DoS) and various spoofing attacks. By establishing a reconstruction error threshold at the 95th percentile (0.7547) of benign sample errors, the framework effectively identifies deviations from normal traffic patterns. Experimental results validate the framework's efficacy in a realistic automotive environment, particularly in detecting previously unseen attack patterns. This research contributes to the advancement of intrusion detection systems for vehicular networks by providing a lightweight, adaptive solution capable of operating within the constraints of automotive embedded systems while maintaining high detection accuracy.