A Hybrid-Based Feature Selection Approach for IDS

Abstract

An intrusion detection (ID) technique classifies the incoming network traffic, represented as a feature vector, into anomalous or normal traffic by a classification method. In practice, it has been observed that the high dimensionality of the feature vector degrades classification performance. To reduce the dimensionality, without compromising the performance, a new hybrid feature selection method has been introduced and its performance is measured on KDD Cup’99 dataset by the classifiers Naïve Bayes and C4.5. Three sets of experiments have been conducted using full feature set, reduced sets of features obtained using four well known feature selection methods as Correlation-based Feature Selection (CFS), Consistency-based Feature Selection (CON), Information Gain (IG), Gain Ratio (GR) and the proposed method on the said dataset and classifiers. In first experiment, classifier Naïve Bayes and C4.5 yielded classification accuracy 97.5 % and 99.8 % respectively. In second set of experiments, the best performance (accuracy) of these classifiers was achieved as 99.1 % and 99.8 % by the method IG. In third experiment, six features are obtained using proposed method and noted the same as 99.4 % and 99.9 %. The proposed hybrid feature selection method outperformed earlier mentioned methods on various metrics.