Identifying P2P Applications Using Ensemble Learning and Co-Training

Abstract

As the evolution of P2P application, the effectiveness of traditional traffic classification approaches for P2Ps has been greatly diminished. Machine learning methods have become popular in this field. In this paper, a machine learning based P2P identification model is proposed, which combines ensemble learning with semi-supervised cotraining techniques, using port-independent and payloadindependent attributes. Ensemble learning can increase the flow accuracy while co-training makes full use of unlabeled samples. In the latter part of this paper, the model is used to identify hiding P2P traffics. A reasonable number of the formerly other samples are identified as P2P flows. Compared with traditional approaches, the effectiveness of our model is proved through a series of experiments from different aspects, including precision, recall and identification rate.