The Study Of P2P Traffic Real-time Identification Based On Support Vector Machine

P2P system broke the traditional Client/Server network architecture, which brought great conveniences to users, while the sustaining growth of P2P applications also affected users' normal actions.P2P traffic must be identified first for controlling the P2P application. Tradition model of P2P traffic identification based on payload features was too monotonous to make adjustments according to network. With encrytion technology been brought into P2P system, and consideration of security and privacy, method based on payload features has lost its efficacy gradually. Although P2P traffic identification based on SVM was able to adjust itself to network changes, playing the advantage of machine-learning, its' training speed was too slow, thus low efficiency to meet the requirements of real-time.According to the above, the thesis firstly proposed an SVM method based on entropy optimization. It built an entropy model by traning space location of sample points, and filtered sample points according to entropy features, which culled out redundant unrelated sample points, while keeping crucial support vectors which decided the end results. By reducing the sample sets, training time was cut off and training became more efficiency. The new SVM model was simple and more efficiency, which surmounted the weak points such as training slowly and low efficiency. By experiments, training time was decreased less than ten percents of original one and it was less than one minute when the sample sets was more than ten thousand, which meet the requirements of P2P traffic real-time identification. Then, the thesis designed a real-time P2P traffic monitoring system based on entropy optimization SVM according to network traffic management demands, realized the system based on campus network architecure and validated its feasibility in real-time P2P traffic identification by experiments.Compared with congeneric methods, the latter method was proved through experiments to be more simple, easy-to-achieve and efficient in sovling linear-divisable classification problems based on two classes, meeting requirements of real-time.