SELECTING KDD FEATURES AND USING RANDOM CLASSIFICATION TREE FOR PREDICTING ATTACKS

The purpose of this study is to identify some higher-level KDD features, and to train the resulting set with an appropriate machine learning technique, in order to classify and predict attacks. To achieve that, a two-steps approach is proposed. Firstly, the Fisher’s ANOVA technique was used to deduce the important features. Secondly, 4 types of classification trees: ID3, C4.5, classification and regression tree (CART), and random tree (RnDT), were tested to classify and detect attacks. According to our tests, the RndT leads to the better results. That is why we will present here the classification and prediction results of this technique in details. Some of the remaining results will be used later to make comparisons. We used the KDD’99 data sets to evaluate the considered algorithms. For these evaluations, only the four attack categories’ case was considered. Our simulations show the efficiency of our approach, and show also that it is very competitive with some similar previous works.