We show that the image representations in a deep neural network (DNN) can be manipulated to mimic those of other natural images, with only minor, imperceptible perturbations to the original image. Previous methods for generating adversarial images focused on image perturbations designed to produce erroneous class labels. Here we instead concentrate on the internal layers of DNN representations, to produce a new class of adversarial images that differs qualitatively from others. While the adversary is perceptually similar to one image, its internal representation appears remarkably similar to a different image, from a different class and bearing little if any apparent similarity to the input. Further, they appear generic and consistent with the space of natural images. This phenomenon demonstrates the possibility to trick a DNN to confound almost any image with any other chosen image, and raises questions about DNN representations, as well as the properties of natural images themselves.
INTRODUCTIONRecent papers have shown that deep neural networks (DNNs) for image classification can be fooled, often using relatively simple methods to generate so-called adversarial images (
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.