On efficiency enhancement of the correlation-based feature selection for intrusion detection systems

The dramatic increase in the network traffic data has become a major concern in security systems. Intrusion detection systems (TDSs), as common widely used security systems for communication networks, are not an exception. An IDS monitors the network traffic to detect attacks through classifying the network traffic data into normal and abnormal classes. Due to the high dimensionality of the network traffic data, it is not always feasible for an IDS to detect intrusions quickly and accurately. Feature selection emerges as a necessary step in designing an IDS to overcome its shortcoming and enhance its performance through the reduction of its complexity and acceleration of the detection process. To this end, in this paper, we address the problem of dimensionality reduction by proposing an efficient feature selection algorithm that considers the correlation between a subset of features and the behavior class label. Correlation-based feature selection (CFS) and symmetrical uncertainty (SU) are the two correlation metrics used to measure the dependency level between features and class labels, and among features. Experimental results on NSL-KDD dataset shows that the proposed approach with fewer features, significantly outperforms the existing schemes in terms of the training time, time taken to build the model, while it preserves or increases the system accuracy. In addition, the efficiency of the proposed feature selection technique is tested on different classification algorithms and comparison results indicates that J48 classifier with the highest accuracy and precision values and lowest miss rate and false alarm rate values, performs better with the proposed feature selection technique.