A Novel Adversarial Defense by Refocusing on Critical Areas and Strengthening Object Contours

The success of deep learning is largely attributed to its representational capabilities, especially in computer vision tasks. However, recent researches have shown that deep neural networks (DNNs) are always vulnerable to adversarial attacks. In order to identify the commonalities between various attacks, we compare the variation between clean and adversarial examples through model-hiding feature visualization methods (i.e., heatmaps), as adversarial perturbations are usually undetectable to the human visual system. It is observed that adversarial examples generated by various attack methods are able to deceive DNNs by distracting critical regions of the image and blurring object contours. Motivated by these findings, we created a straightforward and effective defense method called “Refocusing Critical Regions and Strengthening Object Contours”. This is a defense method based on pixel-attention weights, consisting of pixel channel attention and pixel plane attention. Pixel channel attention reconstructs the critical regions of an image, while pixel plane attention reinforces object contours. Our method demonstrates the effectiveness against different attacks on scalable models and datasets. Moreover, current state-of-the-art defense methods are improved when cascaded with our method.