Resrarch On Internet Traffic Classification Using DBN

In the view of existing models’accuracy is low for P2P traffic, this thesis puts forward a deep learning structure, a semi-supervised learning method, DBN, which is short for Deep Belief Networks as Internet traffic classification method. It constructs an appropriate feature space, establishes traffic classification model, selects the number of hidden units and the number of hidden layers, and improves the accuracy of P2P traffic.This thesis constructs a private dataset of P2P traffic based on the process method. This thesis takes the public dataset provided by LiWei from Moore laboratory, the University of Cambridge and the private dataset as experimental datasets. Then, the thesis respectively takes use of BP and DBN to model, test and analyze the experimental results. It is concluded that the accuracy of P2P is highest with 1-2 hidden layers for BP neural network. But DBN method has the highest accuracy of P2P application with 3-4 hidden layers. For public datasets, DBN method improves F-measure of P2P application 23.3% than BP. For private datasets, DBN method improves the mean F-measure of iQiyiPPS、 Sohu、PPTV and Baofeng 13.2% than BP.