How Can We Deal With Adversarial Examples?

Recently, deep neural network has been widely used in computer vision, machine translation and Go playing and has become the important part of artificial intelligence applications. However, the existence of adversarial examples designed by adversaries could let deep neural networks make incorrect outputs, and fail to work in their applications. Indeed, a considerable amount of work on generating adversarial examples has emerged. Thus, it’s important to investigate adversarial examples in deep learning community. Throughout this paper, we introduce some basic knowledge about attacks on DNNs. Then, we discuss several attack methods to generate adversarial examples. Subsequently, we introduce some defense mechanisms to classify or detect adversarial examples correctly. Finally, we review the challenges in this field, along with the corresponding future directions towards dealing with adversarial examples.