Intrusion Detection by Artificial Neural Networks

Abstract

This paper presents a new approach to intrusion detection using methods of artificial intelligence. Neural networks are suitable for use in intrusion detection systems. To analyze the suitability of using neural networks several data sets were created. They consist of a set of legitimate and malicious communications represented by equally represented samples of data streams, with the number of parameters used varying according to the input parameter optimization method used. For training of the neural networks were used 3 training algorithms: Levenberg–Marquardt algorithm, Bayesian regularization, and scaled conjugate gradient backpropagation algorithm. Dimensionality reduction can decrease the number of features to decrease computational complexity. Two methods are analyzed in the paper: principal component analysis and the stepwise selection method. These methods are compared with results achieved from the training of neural networks for a full set of parameters of the input data sets. The proposed topology of the artificial neural network obtains the probability of correct classification from 80.8 to 84.6% for selected test sets.