Despite their great success in image recognition tasks, deep neural networks
(DNNs) have been observed to be susceptible to universal adversarial
perturbations (UAPs) which perturb all input samples with a single perturbation
vector. However, UAPs often struggle in transferring across