Improved application of transfer learning in network traffic classification

Abstract

When using machine learning for traffic classification, there is such an assumption: the training data and the test data are independently and identically distributed. However, in reality, the assumption that the flow characteristics obey the same distribution may no longer hold because of conceptual drift or regional changes. Existing models will not be able to effectively classify new traffic. The transfer learning method TrAdaBoost has achieved great success in traffic classification and other aspects, but there are some problems, such as too much attention to the difficult-to-classify instances in the target domain, and failure to consider the wrong-classified instances in the source domain. In this study, the method of introducing weight correction factors in TrAdaBoost is used to make the iteration of weights more reasonable, and the effectiveness of this method is proved through theoretical analysis and experiments.