A method for real-time peer-to-peer traffic classification based on C4.5

Abstract

Classification results of network traffic using machine learning rely on attribute information captured at the end of a flow. In contrast, real network requires classifying traffic before a flow has finished. This implies that classification must be achieved using information extracted from the most recent N packets at any arbitrary point in a flow's lifetime. In order to classify peer-to-peer (P2P) applications as early as possible, different P2P applications' characteristics are studied and an attribute set, being able to effectively and promptly distinguish different P2P applications, is proposed. The simulative results using C4.5 decision tree algorithm and sliding window method show that, compared to current attribute sets, this set is more effective in classification, with accuracy achieving 96.7%. Besides, it proves that accuracy using this set keeps stable, though a number of initial packets in a flow are lost.