Research On The Algorithm Of Network Traffic Identification Based On Neural Network

This thesis is supported by the project called Research and Application of Multi-dimensional Sensing Technology over Smart Pipe for Power Information and Communications Network, whose research target is about service-oriented traffic identification and sensing. In this thesis, the algorithm of traffic identification based on neural network is investigated and neural network is introduced to traffic identification so as to improve classification performance.According to the fact that nowadays numerous features are available, feature selection algorithms are studied and then a feature selection algorithm based on the analysis of relationship among attributes named ARAFS is proposed. Based on the theory of neural network, a neural network selection approach of constructing ensemble based on CS algorithm, i.e. CSEN, is put forward. The two methods mentioned above are integrated to verify actual effects of traffic identification on the target datasets collected by Moore and other researchers. The eventual outcome indicates that the proposed approaches manage to identify traffic types at a relatively high rate of accuracy accompanied by robustness as well as good scores on other metrics. The whole thesis can be divided into six parts and its main content is as follows:In the first chapter, research background along with its implications is introduced, basic principle of traffic identification is elaborated and several common techniques are analyzed as well. At the same time, the architecture of this thesis is described.In the second chapter, neural network theory is mainly presented and the drawbacks of BP neural network are pointed out, followed by neural network ensemble. In addition, the generation method of component neural networks and final results in neural network ensemble is summarized.In the third chapter, feature selection algorithms are subjected to analysis in the term of both search strategy and subset evaluation, and thus ARAFS is proposed with tests of its performance on UCI datasets to verify effectiveness.In the fourth chapter, selective neural network ensemble is firstly analyzed, and with the introduction of CS algorithm, CSEN is put forward, followed by verifications about its performance on datasets incorporated by the software of Matlab itself.In the fifth chapter, the two approaches (ARAFS and CSEN) are combined to examine actual effects of traffic identification on Moore dataset with fine metrics.In the last chapter, research work of this thesis is concluded and future direction is also pointed out.