Research Of Intrusion Detection Based On Neural Network Ensemble

With the rapid development and wide application of computer network technologies, computer network security is becoming more and more important. It is an urgent problem in intrusion detection system how to recognize existed attacks and find new attacks rapidly, exactly.Since middle 1980's, the theory and application of neural network have been developing rapidly. Strong learning ability and rapidly answer-finding ability of neural network make neural network become a new solution to the urgent problem of intrusion detection system. The intrusion detection technology based on neural network has been developing rapidly and yielded encouraging effects. Compared to traditional intrusion detection technologies, this technology can detect intrusion rapidly and find many new attacks, but the detection accuracy should be better. So we present a new neural network ensemble method and then apply it to construct the classifier of intrusion detection system.The new neural network ensemble method includes two algorithms, one is producing algorithm of individual neural network based on classified-learning method, and the other one is individual neural network ensemble algorithm based on orthogonal transformation.During the process of producing individual neural network by classified-learning method, we take the component characters of intrusion data into account sufficiently, and split the data into two sets, and then select some samples from the two sets severally in terms of respective selecting policy, and then compose those selected samples into a training subset,and then train a neural network with this training subset. So, we make balance between learning of intrusion samples and learning of normal samples, and improve the detection rate of R2L and U2R that are less-appeared and very dangerous intrusion types.During individual neural network ensemble algorithm based on orthogonal transformation, we found linear relationship between ensemble generalization error and correlation of individual neural networks; and then translate the problem of optimizing ensemble generalization error into the problem of finding the extremum of function; and then find the extremum of function with some conditions by the character of function, so we can make the optimizing of ensemble generalization error. This method alleviates the influences of those factors that may influence ensemble effect directly during individual neural network ensemble with weighted average, and low the complexity of ensemble, and improve the whole intrusion detection rate.Lastly, we apply this new method to intrusion detection system, and then test the performance of this system by experiment, the result of experiment shows that the system is good at intrusion detection.