Siamese-Enhanced Multi-View Representation Learning for Robust Attack Detection in the Internet of Vehicles

The rapid development of the Internet of Vehicles (IoV) has made cyberattacks targeting vehicular time-series data a critical security issue. Existing reconstruction-based detection methods mainly rely on global sequence modeling, which limits their effectiveness in highly dynamic vehicular environments and often results in missed detections of stealthy, localized attacks caused by subtle temporal variations. To address this limitation, we propose the Siamese-Enhanced Multi-View Representation Discrepancy (SMRD) framework for robust anomaly detection in vehicular networks. SMRD introduces a joint representation learning strategy that integrates global trend reconstruction with fine-grained local dynamic modeling. Specifically, the global view captures long-term temporal dependencies through sequence reconstruction and generates a reconstruction representation discrepancy. Meanwhile, the local view employs a novel intra-block and inter-block Siamese network to learn contextual relationships among adjacent time points and produce a Siamese representation discrepancy. Furthermore, to reduce representation bias caused by the non-stationary characteristics of IoV streaming data, an adaptive normalization strategy is introduced to stabilize input distributions during modeling and preserve essential statistical properties after reconstruction. By effectively combining the discrepancies from both views, SMRD enables comprehensive anomaly evaluation at each time point. Extensive experiments on the F2MD simulation platform and multiple real-world datasets, including SPMD, demonstrate that SMRD achieves state-of-the-art performance. Specifically, it achieves the highest F1-scores across all four simulated attack scenarios and delivers superior or competitive performance on public benchmarks, confirming its robustness for securing vehicular networks.