A Study on Detection and Classification of Semi-Supervised Malicious Encrypted Traffic with Variational Self-Encoder Based on Null Convolution

With the rapid development of Internet technology, network traffic shows a significant growth trend in terms of quantity and type. In this context, the categorization and monitoring of network traffic has become a key link in improving service quality, strengthening network management and safeguarding information security. However, with the improvement of users' privacy awareness, more and more applications start to use encryption for transmission, which brings great challenges to the traditional network traffic detection. In response to this technical bottleneck, encrypted traffic detection methods based on deep learning have become a research hotspot, which are mainly based on the statistical features of traffic or directly processing raw traffic data. In this paper, we focus on the challenge of malicious encrypted traffic detection, and propose a semi-supervised detection method with a variational self-encoder fused with null convolution for its scarcity of labeled data. The method realizes semi-supervised encrypted traffic detection and classification by combining the unsupervised model variational self-encoder with the supervised features of multilayer perceptron, and in this paper, the regular convolution in the variational self-encoder is modified to the hollow convolution, which makes the model's sensory wildness improved and the contextual understanding enhanced. The proposed method is demonstrated in the USTC-TFC dataset. Experimental results show that for malicious encrypted traffic classification employing only 5% of labeled data for supervised training can lead to an overall model accuracy of 98%. For malicious encrypted traffic detection employing only less than 2% of labeled data for supervised training can make the overall accuracy of the model reach 100%. This result shows that the proposed semi-supervised learning framework has significant generalization value and practical significance in real applications.