Synthesizing robust adversarial visual samples using Deep Image Prior

Researchers propose a method based on Deep Image Prior (DIP) that can synthesize adversarial image samples that are more robust to affine transformations and visually imperceptible. Traditional adversarial perturbations rely on high-frequency noise that is easily disrupted by subsequent digital operations; this method generates perturbations by reconstructing images with a convolutional network under adversarial constraints, and can also be extended to generate local adversarial stickers. The authors demonstrate such robust adversarial samples on a wide range of images and object categories from ImageNet.