Feature Selection and Classification of Intrusions Using Genetic Algorithm and Neural Networks

Intrusion Detection Systems are one of the emerging areas of Information Security research. They can be implemented using Soft computing techniques. This paper, focuses on multi class classification process whose performance can be significantly enhanced by selecting an optimal subset of input features that is used for training in multi layer feed forward network thereby reducing the false alarm rate. A feed forward network called the back propagation network is trained to classify data as being normal or intrusive. Five training functions are used and analysis is done to decide which training function gives an optimal performance. In addition, the selection of a subset will reduce the dimensionality of the data samples and eliminate the redundancy and ambiguity introduced by some attributes. The user classifier can then operate only on the selected features to perform the learning process. Experiments are performed using kddcup99 dataset. The optimality of the obtained feature subset is then tested and a classification rate of 86% is obtained.