Robust Multi-Class Decision Trees

Abstract

As machine learning is getting deployed more and more in security critical applications, the subject of robustness of machine learning algorithms, i.e. how vulnerable a given algorithm is to inputs that are designed to give a wrong prediction, is getting increased attention from researchers. Early studies focused on neural networks, but decision trees, being another widely used kind of classifier, have also started to receive attention in this area. In recent years, several novel algorithms for training robust decision trees have been proposed. In this study, the comparative performance of two of these algorithms is evaluated. Specifically, the first algorithm, Treant, is extended to support multi-class data, and it is compared to the second algorithm, Robust Decision Trees. The results indicate improved robustness over a naturally trained tree for both algorithms, with Treant being more robust.