Intelligent Cyberattack Detection in the Internet of Vehicles (IoV) Using Hybrid Deep Learning Techniques

Abstract

Rapid growth in the IoV domain has led to significant security risks, making it imperative to use state‐of‐the‐art intrusion detection systems. Conventional methods usually lack the ability to effectively learn spatial, temporal, and global properties from the vehicular traffic data streams. Therefore, this study introduces a novel IDS named FusionIDS. It is different from previous convolutional neural network long short‐term memory (CNN‐LSTM) approaches in that this study considers the integration of Transformer‐based global attention along with SHAP explanations that are specific to IoV IDS. FusionIDS consists of three components, namely, a 1‐dimensional convolutional neural network (1D‐CNN) for extracting local features, an LSTM network for learning sequential information, and a Transformer encoder for global information extraction. In addition, it uses SMOTE to manage the problem of severe class imbalance and SHapley Additive exPlanations (SHAP) for explainability. Results show that FusionIDS obtains around 98% accuracy when tested on various attack cases in IoV environments.