부산대학교 도서관

Deep neural networks are known to be vulnerable to the attack of very small perturbation vectors. One recent method, named universal adversarial perturbation (UAP), is able to generate a single image-agnostic single adversarial perturbation vector commonly applicable to all images in a given data set at one time. However, the computational complexity of UAP increases with the size of data set, thus being very slow to work, especially for large datasets. In this paper, we proposed a fast UAP method, which significantly improves the former's efficiency. Particularly, the proposed method generates a universal perturbation in a mini-batch way with respect to a certain deep classifier, based on multiDeepFool, a newly proposed method that computes an adversarial example for a batch of inputs. The experiments show that our fast UAP is more efficient than the previous UAP algorithm with the nearly same fooling ratio.