A three-class heuristics technique: Generating training corpus for Peer-to-Peer traffic classification

Peer-to-Peer (P2P) applications generate more than 50% of Internet traffic and intensively consume network resources. Statistical machine learning approaches have been proposed as a promising way to detect P2P traffic. However, such systems require retraining in a regular basis. Hence, the generation of good quality P2P and non-P2P examples on a regular basis is not trivial. This paper proposes a three-class heuristics technique to provide regular training corpus generation of P2P and non-P2P traces. In the proposed work, three traffic classes are defined instead of two usually used in typical P2P heuristic classifiers. Based on 22 traffic traces downloaded from different shared resources and captured from Universiti Teknologi Malaysia (UTM) campus network between March and June 2010, the proposed system is evaluated. The result shows that adding the third class improve the accuracy from 93% to 98%. This module could provide quality P2P examples with around 2% class noise that can be used to train P2P classifier on a regular basis.