Abstract
Deep neural networks (DNNs) are under threat from adversarial examples. Adversarial detection is a fundamental work for robust DNNs-based service, which distinguishes adversarial images from benign images. Image transformation is one of the most effective approaches to detect adversarial examples. During the last few years, a variety of image transformations have been studied and discussed to design reliable adversarial detectors. In this paper, we systematically review the recent progress on adversarial detection via image transformations with a novel taxonomy. Then we conduct an extensive set of experiments to test the detection performance of image transformations towards the state-of-the-art adversarial attacks. Furthermore, we reveal that the single transformation is not capable of detecting robust adversarial examples, and propose an improved approach by combining multiple image transformations. The results show that the joint approach achieves significant improvement in detection accuracy and recall. We suggest that the joint detector is a more effective tool to detect adversarial examples.
Abstract (translated)
URL
https://arxiv.org/abs/2201.01080