Concept Drift Detection for Network Traffic Classification
Yulei Wu, Jingguo Ge, Tong Li · 2022
Traffic classification has been widely used for network management. However, characteristics of real-world network traffic are constantly changing due to some unforeseen reasons, e.g. delays caused by network congestion, leading to the occurrence of concept drift and severely degrading the performance of the classifier. In addition, online training of classifiers and concept drift detection often require that all samples need to be labeled. However, labeling network traffic is significantly hindered by its nature of streaming applications. In this chapter, to reduce the number of labeled samples, we propose a concept drift detector based on conditional variational autoencoder ( CVAE ) under the realistic assumption of limited access to the labeled network traffic. Through a large number of experiments, we prove the efficiency of our proposed solution.