Research On Neural Network Classifiers Combination And The Principal Components Analysis Of Intrusion Detection

With the fast-breaking development of Internet, network security has become more and more important in modern computer systems. Due to the frequent security problems, Intrusion Detection has attracted more and more attention as the surveillance of network. However, the detection of intrusion actually is the recognition of intrusion, which belongs to Pattern Recognition category. In this point, classifier designing seems to be un-neglected, and the direction shows that it's popular to use multiple classifiers combination techniques in Intrusion Detection to provide better performance. One of the most notable approaches to learn classifiers combination is the Hierarchical framework. In this paper, the key problem and interrelated technique of Intrusion Detection System are discussed and explored within the application of the Hierarchical combined Neural Network classifiers as follow:To the first point, we've designed a Hierarchical classifiers-combined framework according to the characteristics of BP NN classifier and LVQ NN classifier. This combined classifier can resolve the problem that some classifier can't learn and adapt the new attacks using the independence and the self-adaptation of BP NN. And the competition of LVQ NN helps us to group the unknown data into categories designed by the user. We compare the single BP NN classifier to the Hierarchical combined classifiers on several performances, and the experiment data comes form UCI machine learning data sets.Secondly, considering the high dimension data in Intrusion Detection data sets we used, the need of feature extracting will be put forward indeed. In this paper, Principal Components Analysis is brought forward based on the MATLAB statistics toolbox. And then, we use the 2-level Hierarchical BP-LVQ NN combined classifier for Intrusion Detection, achieving better performance by Resilient Backpropagation training algorithm instead of Levenberg-Marquardt training algorithm.Thirdly, making judgment upon the reliability value by some selection strategy undoubtedly would be ill considered according to the doubtful facts. In this paper, the output of the LVQ NN combined classifier is the vector that represents the reliability the data belong to some category. Thus people can have many choices to make proper judgment.In conclusion, research in this paper has provided theoretical and practical foundation to implement combined Neural Networks classifier and to improve theefficiency of Intrusion Detection data recognition. This research will be a scientific reference for Neural Network classifier designed for Pattern Recognition.