Soft Marking Scheme of SVM Hierarchical Classifiers for Attack Classification

Abstract

Classification is a predictive modelling problem that involves assigning a class label to an instance correctly. There exist several strategies in machine learning to deal with the multi-class classification problems for attack detection. One of the popular strategies is the one-vs-one that decomposes the multi-class problem into multiple binary ones. The approach has been applied in many popular supervised learning algorithms, such as support vector machines. A possible problem of the standard multi-class classification problem is that it lacks correlation between different classes, which can increase overfitting problems and hinder generalization performance. Thus, a possible solution to the problem is to use a hierarchical classification that captures the relationship between classes by dividing the multi-class classification problem into a tree. However, one possible challenge in this approach is selecting parent and child nodes of the tree. The selected nodes should be informative to recognize and then classify different attack classes. One way is by looking at specific domain knowledge to train and build classifiers of the base learners for effective prediction. Thus, a soft marking scheme is introduced to assess a set of binary classifiers to ensure the best overall predictive base learners. Finally, we validate and compare the proposed approach to the standard NSL-KDD dataset. The results show that the proposed method outperforms the standard classifier on the intrusion attack classification.