Universal Adversarial Perturbation Generated by Attacking Layer-wise Relevance Propagation

The vulnerability of Deep Neural Networks (DNNs) to adversarial attacks has become an important research area of machine learning. It has been known that many state-of-the-art DNNs suffer the risk of universal adversarial perturbations, which are image-agnostic and able to lead misclassifications with high probability. In this paper, we propose a novel method to create such universal adversarial perturbations. Our approach is the first to generate universal perturbations by attacking the attention heat maps with the interpretation method, Layer-wise Relevance Propagation. It is demonstrated that our method achieves high fooling ratios on image classification DNNs pre-trained by ImageNet dataset. Moreover, our attack shows good transferability across different DNNs.