Abstract: Recent studies have demonstrated that the deep neural networks (DNNs) are vulnerable to carefully-crafted perturbations added to a legitimate input image. Such perturbed images are called ...