A lightweight anomaly detection model for network traffic using multi scale spatio temporal residual learning

Abstract

With the continuous upgrading of network attacks, network abnormal traffic detection has become a key technology to ensure network security. However, existing detection methods still face challenges in handling large-scale network traffic, solving the category imbalance problem, and improving computational efficiency. To this end, the study proposes an abnormal traffic detection method based on lightweight knowledge transfer anomaly detection network. Firstly, multi-scale residual networks are designed to effectively extract the spatio-temporal characteristics of network traffic. Secondly, on its basis, lightweight knowledge transfer anomaly detection network is proposed in combination with knowledge distillation technique to migrate the knowledge from teacher model to lightweight student model. In the performance test, when the number of iterations was 100, the accuracy of the proposed model was 0.93 and the loss value was 0.27. When the sample size was 5000, the specificity of the proposed model was 0.97, the training time was 22.8 s, and the inference time was 0.06 s. In the simulation test, the traffic loss of the model under high attack intensity was 0.07%. The experimental findings demonstrate that the suggested approach offers notable benefits in handling intricate network traffic data, enhancing the precision and deployment effectiveness of anomalous detection, and holding promise for real-world implementation.