A novel method for intrusion detection in computer networks by identifying multivariate outliers and ReliefF feature selection

The identification of unusual data in computer networks is a critical task for intrusion detection systems. In this study, a novel approach has been proposed for improving intrusion detection system performance by finding multivariate outliers and optimal feature selection. The NSL-KDD dataset consisting of 41 features has been utilized to create and test the system. Firstly, the ReliefF Feature Selection approach has been employed to identify the best features that maintain the classification performance at a high level and 20 features have been determined. Then, to find outliers in the dataset, the Mahalanobis Distance and Chi-Square approaches have been applied. After that, various machine learning methods have been applied to the dataset, and the results have been compared. According to the results, higher classification success has been reached in nearly half the time as a consequence of 20 features obtained from the feature selection and outlier identification processes, compared to the classification done using 41 features. With 99.2187% accuracy, the Random Forest Algorithm has achieved the best classification success. Finally, it has been observed that the suggested approach provides statistically significant results with a quick detection time and higher classification accuracy.