A Hybrid Deep Learning Framework for Intrusion Detection in Automotive Controller Area Networks

The increasing connectivity of modern vehicles has made their internal networks, particularly the Controller Area Network (CAN), vulnerable to a wide range of cyber-attacks (Barnard et al., 2022). Traditional Intrusion Detection Systems (IDS) often struggle with the sophisticated, time-sensitive nature of these threats (Hariharan et al., 2022). This paper proposes a novel hybrid intrusion detection framework that leverages the complementary strengths of Machine Learning (ML) and Deep Learning (DL) to achieve high accuracy and real-time performance. Our methodology involves a multi-stage process: Acquiring the real-world Car-Hacking dataset, preprocessing and balancing data using SMOTE, and innovatively transforming tabular CAN data into a 2D image format. This transformation enables the application of pre-trained, lightweight Convolutional Neural Networks (CNNs) like EfficientNetB0 and InceptionV3 for spatial feature extraction. We further introduce a hybrid model that integrates EfficientNetB0 with a Long Short-Term Memory (LSTM) network to capture both spatial and temporal dependencies in the data, a critical capability for detecting multi-stage attacks (Arreche et al., 2024a). For comparative analysis, traditional ML models such as Logistic Regression, Random Forest, and a hybrid RF-SVM model are implemented. The proposed framework is designed to be computationally efficient, making it suitable for deployment in resource-constrained automotive systems (Musleh et al., 2023). This study outlines a complete, replicable methodology for developing a robust, hybrid IDS capable of defending nextgeneration intelligent vehicles (Awajan, 2023).

View article page

Download PDF