Searching the Adversarial Example in the Decision Boundary

Abstract

Deep learning technology achieves state of the art result in many computer vision missions. However, some researchers point out that current widely used deep learning architectures are vulnerable to adversarial examples. Adversarial examples are inputs generated by applying small and often imperceptible perturbation to examples in the dataset, such that the perturbed examples can degrade the performance of the deep learning architecture.In the paper, we propose a novel adversarial examples generation method. Adversarial examples generated using this method can have small perturbation and have more diversity compare to adversarial examples generated by other method.