| P2P(peer-to-peer) technology plays a crucial role in the network, which has turned into the focus of attention of computer network these years. At the same time P2P technology brings negative effects such as bandwidth occupying, copyright dissension problem and network security problem. In order to improve the healthy development of P2P network, the urgent affair is monitoring the P2P flows in order to reduce the negative effects, and improving the accuracy of the P2P flow detection efficiently has turned into a challenging problem.This paper proposes a P2P flow detection system applying SVM technology, aims at improving the low accuracy for small flows, pretended P2P flows and new appearance P2P flows. This paper makes a research on these fields:(1) SGS frequency conversion packet sampling policy is proposed from the paper to found a balanceable sample between large and small flows which could change the sample frequency flexibly according to the network situation; (2)The paper makes a research on the principle of sample constitution, statistics of 200 flow features according to the balance of different classes of samples and typical sample principle, and finally fix 5 flow features as the sample features on the balance of training efficiency and accuracy of the classifier; (3) A new SVM training algorithm which integrate the incremental training algorithm to iterative training is proposed in the paper, using which integrate the iterative samples and support vectors as the new train sample and make a grouping training to decrease the dimension the of sample training, reduce time complexity and control the sensitivity of SVM classifier.At the end of the paper, we survey the accuracy, failed report percentage and error report percentage of the flow classifier using Andew Moore dataset and dataset from an Internet server, makes a comparison with the flow classifier using traditional SVM method, BP Neural Networks method, Bayes method, Decision-tree model method. |
PDF Full Text Request