The Research Of P2P Traffic Classification Based On Machine Learning Algorithms

More and more P2P applications consume network bandwidth and generate network congestion. The traditional P2P traffic classification methods based on port and payload have many objections. According to the five-tuple definition (source IP address, source port number, destination IP address, destination port number and IP protocol), the packets are classified into bidirectional TCP or UDP flow. The flow attributes irrelevant to protocol and ports are extracted to be characteristic vector that is used to represent the traffic. The traffic classification type is the protocol type of P2P.The traffic characteristic subset is decided by using the ReliefF-CFS method, which combines with ReliefF and CFS to generate characteristic subsets for classifying the P2P traffic. The candidate characteristics are first ordered by the dependency using the ReliefF method, of which larger than threshold is set into the original set for CFS method, then the final optimal characteristic set is decided by using CFS and forward search.The P2P traffic classifier is constructed using C4.5 decision tree, SVM and KNN. And the optimal parameters are determined by using the original characteristic set and estimated with the classification accuracy and time. Part of original packets is statistical to classify P2P traffic. In experiment, the number of 50, 100,150 and 200 bidirectional packets is statistical. The result shows that the promoted method has lower complexity and classification time, and higher classification accuracy.The online real-time classification of P2P traffic is researched by constructing hardware and software platform, which can promote the application.The P2P traffic classification prototype based on machine learning method is designed and implemented, which contains the two stages of construction and classification. Each stage contains data collection, data parse, traffic compound and classification.