A Classification Method Of Network Traffic Based On Universal Features Using Decision Tree Algorithm

Network traffic classification refers to a correspondence between all network traffic and the applications they are generated. It is a precondition of managing and studying the matter of network and plays an important role of the management of QoS and security.Nowadays, most features need a redundant training for different network environment and machine learning algorithm and do not have the ability for real-time classification. In this thesis, universal features that make use of the pattern of communication, payload size and the information entropy are proposed. Experiments show the features have the ability to adapt to various network environments. Machine learning algorithm based on these features can classify the network traffic in short time and can meet the demand of real-time classification.Decision tree algorithm is a typical classification method and widely used in the area of network traffic classification. To avoid the disadvantage of traditional decision tree algorithms, a decision tree algorithm based on the shortest partition distance is proposed. It is more stable and the size of the decision tree is smaller. Classification with this algorithm will be faster than with traditional ones.Besides, with the effect of continuous attributes and noisy data, decision trees may have many branches. It makes the size of the tree larger and may slow down in classification. Therefore, a optimization of the model is done by pruning and discretization of continuous attributes.