Extracting salient features for network intrusion detection using machine learning methods

摘要

This work presents a data preprocessing and feature selection framework to support data mining and network security experts in minimal feature set selection of intrusion detection data. This process is supported by detailed visualisation and examination of class distributions. Distribution histograms, scatter plots and information gain are presented as supportive feature reduction tools. The feature reduction process applied is based on decision tree pruning and backward elimination. This paper starts with an analysis of the KDD Cup '99 datasets and their potential for feature reduction. The dataset consists of connection records with 41 features whose relevance for intrusion detection are not clear. All traffic is either classified `normal' or into the four attack types denial-of-service, network probe, remote-to-local or user-to-root. Using our custom feature selection process, we show how we can significantly reduce the number features in the dataset to a few salient features. We conclude by presenting minimal sets with 4--8 salient features for two-class and multi-class categorisation for detecting intrusions, as well as for the detection of individual attack classes; the performance using a static classifier compares favourably to the performance using all features available. The suggested process is of general nature and can be applied to any similar dataset.