Research On Network Traffic Classification Based On Neural Network

In order to better satisfy more and more fine demand of users for various Internet service quality, the traffic classification is an important link of the network management process. There is the current situation that the traditional classification methods existing problems such as low accuracy and limited application region could not meet the pragmatic demands of network manager, which the methods are based on default matched port numbers and signatures analysis. Studying new traffic classification method has important practical significance. Aiming at the different flow samples and the different applied aspects, this dissertation focuses on introducing Neural Network (NN) to traffic classification. The main contents of our study include:Supervised BP (Back-Propagation Algorithm,BP) algorithm which has the high detection precision is introduced. Because of the problems in current work that relies on the standard BP algorithm, such as slow convergence and local minimal, an effective approach for network traffic classification named GA-LM is proposed. This method applies L-M (Levenberg-Marquardt) algorithm which of improved BP algorithms and Genetic Algorithm (GA) that optimizes neural network weights. Experimental results using labeled traffic dataset indicate that this traffic classification method will speed up the convergence of the neural network and improve the classification performance, which is particularly suited for traffic management and traffic accounting.SOM (Self -Organizing Mapping,SOM) method that is a characteristic of strong self-organization and adaptability is improved. Aiming at unsupervised SOM has the lower detection precision than supervised methods, a network traffic classification method based on Supervised Self-Organizing Maps (SSOM) is proposed. The theorys and the experiments proved that this approach is superior to the traditional SOM in aspect of classification accuracy. Moreover, network application types can be analyzed easily according to the result of classification, which makes the proposed method extremely useful for network traffic analysis. However, due to the supervised algorithm requries labeled training samples and network traffic is highly nonlinear, a network traffic classification method named Kernel-SOM (KSOM) is proposed based the introduction of the concept of kernel-learning. Simulation results show that this method, which the performance of its output mapping is better than classical SOM, can identifies the unknown application type of flow samples and achieves high classify accuracy. So it applies to the research of network application. Finally, the network traffic classification model based on KSOM is realized. This model possesses the functions of which capturing packet in online phase and forming into traffic samples in offline phase, and employing proposed KSOM method in classification. After the testing under the real network environment, this model is demonstrated that it has well classification performance, and is a visualized way to display output results which makes it clear, easy to analysis traffic.