A Hybrid Deep Learning Framework for Multi-Modal Intrusion Detection in Internet of Vehicles

The increasing complexity of in-vehicle networks has heightened the vulnerability of modern automobiles to cyber-attacks, necessitating robust and real-time intrusion detection mechanisms. In this study, we propose a hybrid deep learning architecture that integrates 1D CNN-LSTM, Wavelet-based Vision Transformers (ViT), and Graph Attention Networks (GAT) to effectively model the multimodal nature of vehicular communication data—comprising binary payloads, normalized sensor readings, and hexadecimal metadata. A gated attention fusion mechanism dynamically learns the relevance of each modality during inference, enhancing classification performance. Evaluated on the CICIoV2024 dataset, the proposed model achieved a ROC-AUC of 99.78%, an Attack Detection Rate (ADR) of 98.9%, and reduced the False Positive Rate (FPR) to 1.2%, significantly outperforming classical machine learning baselines. Furthermore, the model demonstrated real-time inference capability with an average latency of 21.3 ms on an NVIDIA GeFORCE 3070. These results highlight the effectiveness of modality-aware fusion and deep multimodal feature extraction in securing in-vehicle networks against diverse cyber threats.