Feature selection, learning metrics and dimension reduction in training and classification processes in intrusion detection systems

Abstract

This research presents an IDS prototype in Matlab that assess network traffic connections contained in the NSL-KDD dataset, comparing feature selection techniques available in FEAST toolbox, refining prior results applying dimension reduction technique ISOMAP. The classification process used a supervised learning technique called Support Vector Machines (SVM). The comparative analysis related to detection rates by attack category are conclusive that MRMR+PCA+SVM (selection, reduction and classification techniques) combined obtained more promising results, just using 5 of 41 available features in the dataset. The results obtained were: 85.42% normal traffic, 80.77% DoS, 90.41% Probe, 91.78% U2R and 83.25% R2L.