Research On P2P Traffic Classification Based On Time-Frequency Parameters

With the development of the Internet, people need higher quality Internet servicethan used to be.P2P application traffic accounted for more than half of the networktraffic. The P2P applications bring huge bandwidth pressure to the network, affectingthe normal bandwidth usage of other Internet applications. In order to ensure thenormal operations of the network, the network administrator can identify the P2Papplication traffic by P2P traffic classification techniques. Thus the networkadministrator can limit the bandwidth of P2P applications to ensure the rationalbandwidth usage of other applications.In2005, Moore et al. published an article at SIGMETRICS that using Bayesianclassifier for traffic classification. In that article, they came up with249TCPclassification parameters, which are based on the flow’s time-domain characteristics.However, the traffic classification results based on time-domain characteristicparameters are unstable. So, in this thesis, we did the following work:1.The main innovation of this thesis is time-frequency flow characteristicparameters for P2P traffic classification.This thesis first analyzes the characteristics ofP2P applications communication,and found that compared to the traditional C/Sapplications,the P2P applications have repeated transmission characteristics of the fileblock. In order to get this characteristic, this thesis extracts payload sequence andpacket interval arrival time sequence, and applies time frequency transform to thesesequences. After getting the transformed sequences, we compute time frequency flowparameters.When compared to the time domain to frequency characteristic parameterscan better reflect the characteristics of the P2P application communication, therebyobtaining a more stable classification results.2. Network classification is a long time envolving research area. In order tosupport P2P trafficclassification and other Internet applcication classificationresearch,we developed a network classification platform. This platform supports big fileanalysis. This platform can make the job of parameters extraction much easier.Researchers only need to maintain flow parameters plugins to research network classification. Now, this platform has been used to other traffic analysis area.3. At the experiment part, this thesis first analyze flow bytes distribution of thenetwork traffic, and find that small number of flows takes major network bandwidth.This means the application based on P2P classification should not only pay attention toflow accuracy and also bytes accuracy. Then compare the flow encrypted and the flownot encrypted, we found that whether the P2P flow encrypted or not do not have muchaffection on P2P traffic parameters. This means that use time-frequency parameterscan be used to P2P traffic classification. Finally, in order to compare the time domainparameters, time-frequency parameters and mixed parameters, we use cross sitevalidation. The result turns out that time-frequency parameters are much more stablethan time domain parameters.