Optimal feature selection for machine learning based intrusion detection system by exploiting attribute dependence

Feature Engineering plays an important role in the development of a Machine Learning-based Classifier; especially for Intrusion Detection Systems. It helps in reducing the dimensions of the available datasets, training time, and computation costs; yet improves the performance and detection accuracy of the model. Feature Selection is the most common technique used for reducing the dimensionality of the available dataset. The higher the dimensions of the dataset; the more will be the training time required by the Machine Learning model to process (train and test) the dataset. This paper proposes two approaches for constructing an optimal feature subset, termed Dense_FR and Sparse_FR; to reduce the dimensions of the dataset, based on Kendall's Correlation Coefficient and Mutual Information. Mutual Information is an important and common metric used for Feature Selection. It tries to reduce the amount of uncertainty by incorporating additional attributes. Kendall's Correlation Coefficient is a stricter and consistent correlation coefficient when compared to Pearson's Coefficient or Spearman's Coefficient. The names Dense_FR and Sparse_FR justify the number of features generated in the optimal feature subsets; there are fewer features in the optimal subset generated by the Sparse_FR approach when compared to the Dense_FR approach. Results show that the proposed approaches improve the performance of classification. (c) 2021 Elsevier Ltd. All rights reserved. Selection and peer-review under responsibility of the scientific committee of the Technology Innovation in Mechanical Engineering-2021.