Active learning for P2P traffic identification

P2P traffic identification methods by using machine learning have been provided in a great number of works, which suffer from a large and representative labeled sample set. To overcome the sample labeling problem, a new P2P traffic identification approach by active learning called P2PTIAL is presented. P2PTIAL is composed of two parts: support vector machine as learner and uncertainty selection based on distance. In order to improve the effectiveness of P2PTIAL, we add filtering policy and balanced policy to P2PTIAL. Firstly, we use support vector data description (SVDD) theory to filter some unlabeled samples, which have little contribution on active learning, and so it can save computation cost and storage space. Secondly, we use the unlabeled sample's pre-labeled information to develop balanced policy, which can keep balanced learning. Lastly, we support our design with extensive simulation experiments, and our results show P2PTIAL is feasible.