Research On Neural Network Ensemble And Its Application To The Identification Of P2P Traffic

At present, the application of P2P technique is more and more popular, but it also brings many problems such as big amount of network bandwidth exhaustion, difficulties on management, copyright, security and spamming. It is of great theoretical and realistic importance to study on efficient identification of P2P traffic to improve network service through monitoring and managing. In the monitoring and management system, P2P traffic identification technology also has important meaning.Through the international reports, the method based on neural network to identify P2P traffic can achieve better results. But the system generalization of the method is weak, and this dissertation will solve the problem by neural network ensemble.There are three innovations in this article. Firstly, I construct a neural network ensemble based on adaptive genetic algorithm, which can improve the generalization ability of learning systems. Secondly, I introduce this method into RBF and FUZZY ARTMAP neural network ensemble and make P2P traffic identification by RBF and FUZZY ARTMAP neural network ensemble. Finally, I build a simulation model to test and compare the P2P traffic identification method of RBF and FUZZY ARTMAP neural network ensemble.Firstly, it makes a simulation of P2P traffic identification by BP neural network,RBF neural network and FUZZY ARTMAP neural network in this paper. As a result, three kinds of neural network could recognize the flux of P2P very clearly. As to the cost of training time, RBF neural network is the best and BP neural network is the worst.Secondly, it makes an integrated method by RBF and FUZZY ARTMAP neural network. As a result, on the basis of costing some time on training and distinguishing, these two kinds of neural network ensembles have improved much on recognizing the flux.Finally, it optimizes the neural network ensemble and gets a result as that:the average recognition rate is 99.15%, the average training time is 2.5872s and the average recognition time is 0.9174s of RBF neural network ensemble based on the simple average. We raise the average recognition rate by 1.5%. The average recognition rate is 99.48%, the average training time is 6.2824s and the average recognition time is 1.4193s of FUZZY ARTMAP neural network ensemble based on the simple average. We raise the average recognition rate by 0.6%.By analyzing the simulation results we can see that the RBF neural network ensemble is very bonuses because of that the training time and recognition time is less, and that the average recognition rate is lower than FUZZY ARTMAP neural network ensemble. In the end, I suggest that we can use the RBF neural network ensemble to identify P2P traffic when we just can only use software, and the FUZZY ARTMAP neural network ensemble is better when we can use hardware and parallel computer.