Double Targeted Universal Adversarial Perturbations

Despite their impressive performance, deep neural networks (DNNs) are widely known to be vulnerable adversarial attacks, which makes it challenging for them deployed in security-sensitive applications, such as autonomous driving. Image-dependent perturbations can fool a network one specific image, while universal capable of fooling samples from all classes without selection. We introduce double targeted (DT-UAPs) bridge the gap between instance-discriminative image-dependent and generic perturbations. This perturbation attacks source class sink class, having limited effect on other non-targeted classes, avoiding raising suspicions. Targeting simultaneously, we term attack (DTA). provides an attacker with freedom perform precise DNN model little suspicion. show effectiveness proposed DTA algorithm wide range datasets also demonstrate its potential physical (Code: https://github.com/phibenz/double-targeted-uap.pytorch).