Critical Study of Supervised Learning Techniques in Predicting Attacks

ABSTRACT This paper presents a study about the use of some supervised learning techniques to predict intrusions. The aim of the research is to analyze the performances of such techniques to determine which one best addresses the intrusion detection problem. The performances of six machine learning algorithms involving C4.5, ID3, Classification and Regression Tree (CART), Multinomial Logistic Regression (MLR), Bayesian Networks (BN), and CN2 rule-based algorithm are investigated. The “boosting-arcing” concept was used to obtain a better prediction model while executing a machine learning method. KDD'99 data sets were used to evaluate the considered algorithms. For these evaluations, three cases were considered: the whole attacks case, the five behaviors classes' case, and the two behaviors classes' case. The performances of each technique were compared, and simulations showed that our approach is very competitive with some previous works.