An Optimized Algorithm for Speeding up Universal Adversarial Perturbation Generation

Abstract

Universal adversarial perturbation, which is a security vulnerability in convolutional neural networks (CNNs), can fool convolutional neural network models on a set of images by a single perturbation vector. One recent algorithm, named UAP, generates universal adversarial perturbation iteratively by aggregating the smallest adversarial perturbations with respect to each image, but it ignored the orientations of perturbation vectors; consequently, the magnitude of the universal adversarial perturbation cannot efficiently increase at each iteration, thereby resulting in slow universal adversarial perturbation generation. Hence, to expedite the generation of universal adversarial perturbation, we propose an optimized algorithm to generate universal adversarial perturbation based on the orientations of perturbation vectors and aggregate adversarial perturbations with similar orientations. The proposed algorithm is compared with the original algorithm on ImageNet dataset, experimental results show that our proposed algorithm is more efficient and can reduce the number of training images compared with the UAP with nearly the same fooling rate.