Application of Partial-Connected Dynamic and GA-Optimized Neural Networks to Misuse Detection Using Categorized and Ranked Input Features

The number of attacks in computer networks has grown extensively, and many new intrusive methods have appeared. Intrusion detection is known as an effective method to secure the information and communication systems. In this paper, the performance of Elman and partial-connected dynamic neural network (PCDNN) architectures are investigated for misuse detection in computer networks. To select the most significant features, logistic regression is also used to rank the input features of mentioned neural networks (NNs) based on the Chi-square values for different selected subsets in this work. In addition, genetic algorithm (GA) is used as an optimization search scheme to determine the sub-optimal architecture of investigated NNs with selected input features. International knowledge discovery and data mining group (KDD) dataset is used for training and test of the mentioned models in this study. The features of KDD data are categorized as basic, content, time-based traffic, and host-based traffic features. Empirical results show that PCDNN with selected input features and categorized input connections offers better detection rate (DR) among the investigated models. The mentioned NN also performs better in terms of cost per example (CPE) when compared to other proposed models in this study. False alarm rate (FAR) of the PCDNN with selected input features and categorized input connections is better than other proposed models, as well.