Research On Adaptive Traffic Identification Based On Concept Drifting Detection

Along with the development of network and the improvement of hardware, thetechnology of traffic identification plays a more important role. Because of thecomplication of network application and personal privacy, some identificationtechnology, such as DPI, already can’t meet the needs of reality. The trafficidentification based on machine learning has been becoming the current mainresearch content. Meanwhile, the concept drifting seems more and more urgent.Firstly, the basic principle of concept drift was expounded in detail, from theperspective of network traffic. The related definition of concept drift and conceptdrift detection has involved, laid a solid foundation for later study.Secondly, the concept of drift detection algorithm based on error rate, in thefield of data stream mining, is analyzed. Its limitations limit the range of application.Through the analysis of the detection principle of concept drift, chi-square test andFisher test, concept drift algorithm based on statistical theory is proposed. Further,detailed theoretical demonstration and experiment data verify the validity of thealgorithm.In addition, considering the imbalanced category of the real networkenvironment, we assess respective characteristic of three machine learningalgorithms, including the bayesian estimation, the decision tree and support vectormachine (SVM). The decision tree algorithm is chosen to further network trafficidentification study.Finally, according to concept drift algorithm based on statistical theory, threedifferent forms of adaptive traffic identifier is designed. Each identifier has beenbuilt on an ensemble learning method. Through the experiment, it can be seen thateach identifier has its own advantages, is suitable for different traffic identification.