Applying Machine Learning Combined with Dimensionality Reduction to Enhance Intrusion Detection Systems

An intrusion detection system (IDS) is a system designed to detect potential intrusion activities, such as malicious activities, computer attacks, or the spread of viruses, and then inform the relevant personnel. This study aims to combine different dimensional reduction algorithms and machine learning classifiers to find a more effective approach for network intrusion detection. In the face of a large amount of high-dimensional data, how to effectively utilize and optimize this data to improve detection accuracy is an important research question. Thus, this study investigates the performance of six common machine learning classifiers: Logistic Regression (LR), K-Nearest Neighbor (KNN), Naive Bayes (NB), Random Forest (RF), Decision Tree (DT), and Multilayer perceptron (MLP), with and without applying dimensionality reduction algorithms. The dimensional reduction process is conducted in three different ways, Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and Factor Analysis (FA). The six classifiers mentioned above are applied to these two parts of the data (applying dimensionality reduction algorithms before using machine learning classifiers and not applying dimensional reduction algorithms) respectively. This study mainly focuses on the number of features left after applying dimensional reduction, run time, recall, and $\mathbf{F 1}$ score in order to compare the performance of different models. This paper finds out that the combination of RF and PCA reaches the optimal performance among the models using dimensional reduction algorithms before applying machine learning classifiers. The results also highlight that the dimensional reduction combined with machine learning can optimize intrusion detection systems. Three of the variables are found to be of great significance for detecting an attack.