ENHANCED STOCHASTIC LEARNING FOR FEATURE SELECTION IN INTRUSION CLASSIFICATION

Feature selection is very important in various fields due to the enormous data processing requirements. Many researchers are investigating data reduction and feature selection, particularly in network traffic reduction for network intrusion detection systems. In this paper, we suggest a method that selects the useful features for intrusion classification, decreases the storage of rules and computational time, and increases the classification accuracy. A proposed many-to-many Kullback-Leibler (K-L) divergence is applied to the probability distribution that is calculated by the histogram estimator to select the specific features. This method is an improvement on the previous method that only calculated the distance between the normal and each intrusion. To verify the method, we present experimental results of the classification rates and false positive rates for accuracy, the number of rules generated, and the features selected for accuracy and faster detection. The results of the applying method show that the number of selected features is reduced from 41 to 19. Also, the accuracy rate is increased by 0.1588 pecent, whereas the false positive rate is decreased by 0.0033 percent. Therefore., we confirm the classification accuracy of the proposed method and support its usefulness for data reduction.