Improving the Performance of Intrusion Detection System through Finding the Most Effective Features

Abstract

In recent years, we witnessed the ensuing surge in massive numbers and types of attacks. The future years will continue these trends but at a faster pace as a result of increasing the number of devices and the development of IoT devices. Thus, it becomes really important to detect different types of threats and hence secure these resources. To that end, previous works examined different feature selection techniques and machine learning algorithms. However, they are either suffer from a low detection accuracy or are not able to detect various types of attacks particularly the low-frequency attacks like worms. In this paper, we use multiple feature selection algorithms to find the subset of the more relevant features regarding each type of attack. Forward Selection Ranking and Backward Elimination Ranking algorithms are used along with decision tree classifier and random forest classifier. The system is evaluated in terms of accuracy, precision, sensitivity, and F-score and shows very high performance in detecting all types of attacks. It can detect all types of attacks with an accuracy rate of 99.9% and 99.96% for binary classification.