Real-time feature selection in traffic classification

Classification of network traffic using port-based or payload-based analysis is becoming increasingly difficult since many P2P applications are using port hopping and encryption to avoid detection. As a result, the attention of research has diverted to the application of machine learning (ML) techniques on traffic classification. Feature selection plays an important role in machine learning-based traffic classification scheme. However, most of the feature selection algorithms proposed so far are not effective in online traffic classification, in which the feature subsets selected cannot be used online. In this paper, different types of features used in traffic classification are presented, and meanwhile, the accuracy of different feature selection algorithms is analyzed and compared when they are used in traffic classification, especially in P2P traffic classification. In order to realize the online traffic classification, we propose a real-time feature subset. As the results show, very good performance will be achieved when the real-time feature subset is used in decision tree based ML algorithm. Meanwhile, the build time of real-time feature subset is shorter than that of other feature subsets, which makes it a good candidate for online traffic classification.