Dimensionality Reduction and Supervised Learning for Intrusion Detection

Abstract

This paper proposes the application of supervised machine learning approaches for Intrusion detection. Moreover, it examines the effect of applying dimensionality reduction on classification performance. For validation, we use a subset of the NSL-KDD dataset, a widely applied dataset in this domain. Our results indicate that seven classification algorithms perform well on the dataset and based on the accuracy and false positive rate measure. The best reported accuracy results are using Random Forest algorithm with 80.5% accuracy. To enhance the classification performance, we use two dimensionality reduction algorithms: Principal Component analysis (PCA) feature reduction algorithm and BestFirst feature selection algorithm. PCA is effective in enhancing the performance of four algorithms with ranges of improvements from (1.0% – 4.1 %). Moreover, BestFirst algorithm is effective in enhancing the performance of five algorithms with improvements ranging from (0.1 % – 2.0%). In addition, there is saving in the training time after feature selection with slightly better results compared to the original full feature set.