Edge Computing Intrusion Detection Algorithm Based on Transformer and Autoencoder

Existing intrusion detection methods for the Internet of Vehicles often face challenges in handling high-dimensional data characterized by redundant and highly correlated features, which hinders the ability of the reconstruction model to accurately capture representative features, leading to imprecise data reconstruction and low detection accuracy. To address this issue, an intrusion detection algorithm for edge computing that integrates Transformers and autoencoders is proposed. The proposed approach extracts spatial features using stacked autoencoders and temporal features through variational mode decomposition combined with a multi-head attention mechanism, achieving effective spatiotemporal feature fusion and reducing high-dimensional data to low-dimensional representations. These fused features are then fed into an intrusion detection algorithm that leverages the global learning capability of Transformers and the feature learning ability of autoencoders, improving the detection of both known and unknown attacks. Furthermore, a collaborative framework for Internet of Vehicles intrusion detection in edge computing environments is designed, enabling edge nodes to work together and share resources, thereby improving the ability of the whole system to respond to intrusion. Evaluated on the NSL-KDD and UNSW-NB15 datasets, the algorithm achieves detection accuracies of 97.26% and 98.93%, respectively, with F1-scores of 97.45% and 99.29%, outperforming other SOTA methods. Real-world vehicle experiments further validate the algorithm has high accuracy and strong generalization capabilities, demonstrating its effectiveness in enhancing intrusion detection for the Internet of Vehicles.