The Application Of BP Neural Network In Intrusion Detection System And Its Optimization

BP neural network has many merits for its application in intrusion detecting system , while also having some defects , such as its slow running rate. The commonly used LMBP algorithm can not be applied to the intrusion detecting system for its running speed , even with the quick constringency rate. The LMBP algorithm is optimized by utilizing the KDD99 data set , selecting the suitable data and adding some limiting conditions. The computation results before and after the optimization are compared , which confirm the optimized algorithm is effective. The optimized algorithm apparently increases the speed of BP neural network , and has certain practical significances , when it is applied in intrusion detecting system.This paper is composed of four chapters, which are independent and correlative to one another. In chapter 1, i. e. prologue, the intrusion detecting and basic concept of BP neural network and its development history and the recent applications are introduced concisely. It's the basic tool needed in this paper. In section §1.1, the basic concept of intrusion detecting are introduced. In section §1.2, the basic BP algorithm are introduced. In section §1.3,some improvement methods of BP algorithm are introduced, which include Momentum ,Variable Learning Rate , Conjugate Gradient and Levencage-Marquardt algorithm. Among them, we introduced the Levenberg-Marquardt algorithm with emphasis. This chapter is a basis for the following all charpters.In section §2.1 of chapter 2, the basic situation of KDD99 data set is introduced. There is so much attribute in this data set, in order to simplification the question, we needs to reduce the attribute in the data set. In section §2.2, the Rough set theory which reduce the attribute in KDD99 data set is introduced. In section §2.3, we use the attribute reduction algorithm based on Rough set theory reduce the attribute in KDD99 data set, and get a good result. This indicated that using rough set reduce the attribute is extremely effective. In section §3.1 of chapter 3, some achievements and application of BP neural network in intrusion detection system are introduced firstly. In section §3.2, according to the result in previous chapter, which the data needs to use in KDD99 data set has been made further processing, and we finally determined the experiment data. In section §3.3, we discussed the question of network parameter choice. It is quite difficult to determine the number of hidden layer. Finally we proposed the generalization ability question of BP neural network. In section §3.4, we using the function in Matlab Neural Network Toolbox training and testing the network, by using the definite network parameter. In section §3.5, we give an optimized algorithm. The computation results before and after the optimization are compared, which confirm the optimized algorithm is effective.In chapter 4, we introduce the question of generalization ability question of BP neural network. In section §4.1, we introduce the uncertainty relation when BP neural network is overfitting:By the equation, the result in section §4.2 is obtained . In section §4.3—§4.4, the equation of computing the number of hidden layer is obtained:This is the equation to compute the number of hidden layer in this paper. From the result in the third chapter, we know, the number of network hidden layer determined by the equation make the network has well generalization ability. Finally we point out the problems which will be solved in the field and the plan we will do in the future.