Evaluation of Supervised Machine Learning Algorithms for Multi-class Intrusion Detection Systems

Abstract

The increased demand for network security nowadays is becoming a crucial strategy. Accordingly, a need for intrusion detection system is essential to track cybersecurity attacks. Thus, some protection strategies are necessary for this purpose. However, the current intrusion detection systems are still developing and looking for more accuracy. In this paper, supervised learning algorithms (Random Forest, XGBoost, K-Nearest Neighbors (k = 5), Artificial Neural Network, Logistic Regression, Support Vector Machine, and LASSO-LARS) are trained and tested to a preprocessed dataset. It contains benign and up-to-date common attacks (DoS, Probe, R2L, and U2R). In order to measure the performance of each supervised learning algorithm, the F1-score is calculated. As a result, the random forest, XGBoost, and K-nearest neighbors (k = 5) algorithms have better accuracy than the others, having values of 99.7%, 99.1%, and 97.6% prediction success rate, respectively. The least performance is for logistic regression algorithm with a prediction accuracy rate value of 71.4%.