Study And Implementation Of Network Traffic Classification Base On Machine Learning

Network traffic classification is one of the key supporting techniques in the areas such as network management, traffic engineering and security detection. Its research has great useable value under the circumstance that the kinds of Internet applications are growing rapidly and the variety is changing steadily. As the classification method of machine learning based on traffic statistic features is able to overcome the drawback of traditional port or payload based method, it becomes the research hotspot in the research directions of machine learning application and Internet flow classification.In this thesis, the machine learning method is adopted to experiment and optimize the accuracy on a well known data set and the result is used to direct the design and fulfill a practice system. The main work includes four aspects. First, the standard machine learning model, issue in the flow feature space, basic process of the method, and criteria of accuracy of network flow classification are described. Second, through the experiments on the classical Moore data set, the accuracy value of SVM and improvements by subsample are given. The influence of the training set size and the port information on the accuracy are also studied. Next, relations between variance of feature set and the accuracy is investigated to choose a feature set concerning both the holding of accuracy and the convenience of measuring. In the end, the design and fulfillment are introduced, including the general plan and the key techniques such as traffic capture, flow recomposing, feature measurement, and flow labeling.We tested the traffic classification system by the collection and reorganization of the flows from a campus LAN connected to Internet. It assisted the statistic and prediction of the usage of network effectively.