Research On Identification Methods Of P2p-streaming Media Traffic Based On Incremental Svm Learning

The P2P streaming media applications mushroom on the Internet along with development of the technologies of peer to peer and multimedia, at the same time it brings potential safety hazard in society and public opinion.Because the Internet is an open environment, if them master the technologies of P2P streaming media, hostileforce within and beyond the borders can use it to spread reactionary program, pornography and violence. The group of FaLunGong has build a P2P streaming media application - IPPOTV, them make use of IPPOTV to declare reactionary opinion, if it can't be controlled in time, the influence to the society will be execrable. But the P2P streaming media applications has some characters like P2P applications, for example non-centralization, self-organization and so on, it's hard to supervise them effectively. So it's important to identify the traffic of P2P streaming media application quickly and accurately.Our group has researched on method to identify P2P streaming media traffic. We provide a new approch using packet size distribution character. The approach based on Support Vector Machines (SVM) has excellent performance. But new P2P streaming media applications will come forth ceaselessly in real network. And the existent P2P streaming media applications will change their protocol. The method carries out classification depending on a fixed model which produced by a train process, so the method has not the ability of incremental learning, that is to say, the mothod based on SVM can't adapt the real network very well.In order to let the method suit the complicated circumstance of the network, this paper has research on some technologies as follows:Firstly, this paper provides a method based on packet size combination character. this method has solved the problem of attribute incremental learning.Secondly, because new P2P streaming media applications will appear in the network constantly, this paper proposes a class incremental learning algorithm-CIOOL. CIOOL can make use of the knowledge of old model fully, and can accumulate knowledge rapidly. So this method can identify new P2P streaming media application in a short time. In order to enhence the performance of CIOOL, this paper proposes a sample refine algorithm-FSR. It can reduce the number of samples remarkably, but it keeps the space character of the samples all the same. FSR improves the efficiency of CIOOL remarkbly.Thirdly, because the existent P2P streaming media applications will change protocol or become invalid, this will induce the SVM model produce invalid knowlege. In order to solve this problem, this paper proposes a class replace incremental learning algorithm-CRIOOL. This algorithm can update the knowledge of the model rapidly. Fourthly, based on the technologies this paper researched, a prototyp identified system TV-EYE has been designed, and a sofeware also has been done. Experiment has been done using TV-EYE in a campus network. Results of experiment show that the technologies are effective and TV-EYE has the ability of supervising P2P streaming media.