Information theoretic feature extraction to reduce dimensionality of Genetic Network Programming based intrusion detection model

Intrusion detection techniques require examining high volume of audit records so it is always challenging to extract minimal set of features to reduce dimensionality of the problem while maintaining efficient performance. Previous researchers analyzed Genetic Network Programming framework using all 41 features of KDD cup 99 dataset and found the efficiency of more than 90% at the cost of high dimensionality. We are proposing a new technique for the same framework with low dimensionality using information theoretic approach to select minimal set of features resulting in six attributes and giving the accuracy very close to their result. Feature selection is based on the hypothesis that all features are not at same relevance level with specific class. Simulation results with KDD cup 99 dataset indicates that our solution is giving accurate results as well as minimizing additional overheads.