Anomalous Traffic Detection Using an Enhanced MobileViT Model

In recent years, anomalous network traffic has become an important threat to the cyber security. To address the shortcomings of lightweight networks under complex background conditions, this paper proposes a MobileViT-based enhanced model, which is called MDFA-MobileViT. With a Multi-Dimensional Feature Attention (MDFA) module in the input and the output of the Transformer units, local key feature extraction capacity and global feature modeling capacity is improved. Moreover, 3×3 convolution kernels within Fusion module have been substituted by 1×1 convolutions to efficiently decrease the complexity. In order to further enhance classification ability, an innovative loss function that fusion crossentropy loss, feature center loss and attention-guided loss is proposed. The experimental results reveal that the proposed MDFA-MobileViT model obtains an excellent detection accuracy on the USTC-TFC2016 datasetdemonstrate that the proposed MDFA-MobileViT model obtains an excellent detection accuracy of 99.76%, which is remarkable comparing with the baseline models. Overall, the MDFA-MobileViT model demonstrates strong performance in traffic classification tasks, particularly in detecting complex attack traffic and low-quality traffic samples.