Feature Selection in Machine Learning-Based IDS Performance

Computer security faces many challenges, including the detection of attacks generated by various intrusions. Intrusion detection systems (IDS) have different approaches, and those based on supervised learning have high capabilities to predict different types of attacks. As in other cases, machine learning algorithms supporting these tools benefit from dimensionality reduction. In this paper we present the results of applying different algorithms to obtain subsets of features that maintain good performance in classification tasks and apply the ensemble operations strategy to obtain the final features that are fed to the classifiers to determine the outputs. The results show that in the different classifiers all the metrics go down if the number of features is decreased in high degree, but that with the operation union of the subsets of the features obtained from the importance of the features, the number of attributes obtained is significantly lower and the good performance of the classifier is maintained.