Visual Analysis of Adversarial Examples in Machine Learning

In recent years, machine learning (ML) has received much attention and has been rapidly developed to handle a variety of practical tasks. Among the various ML methods, deep neural networks tend to achieve the best performance thus far. However, adversarial examples (AEs) pose a serious threat to ML models. AEs, which are generated by slightly modifying benign (normal) data, can mislead the prediction of a targeted ML model. In this chapter, current research trends in the visual analysis of AEs are presented. Visualization is a technique that is helpful to intuitively explain and understand complex concepts. This chapter classifies current work into several categories, namely, visualizing the generation of AEs, the properties of AEs, methods of distinguishing AEs, and the robustness of models against AEs. At the end of the chapter, current challenges and interesting future research directions in this field are also discussed.