DCVAE-adv:A Universal Adversarial Example Generation Method for White and Black Box Attacks  

作　　者：Lei Xu Junhai Zhai 

机构地区：[1]College of Mathematics and Information Science,Hebei University,Baoding 071002,China

出　　处：《Tsinghua Science and Technology》2024年第2期430-446,共17页清华大学学报（自然科学版（英文版）

基　　金：supported by the Key R&D Program of Science and Technology Foundation of Hebei Province(No.19210310D);the Natural Science Foundation of Hebei Province(No.F2021201020).

摘　　要：Deep neural network(DNN)has strong representation learning ability,but it is vulnerable and easy to be fooled by adversarial examples.In order to handle the vulnerability of DNN,many methods have been proposed.The general idea of existing methods is to reduce the chance of DNN models being fooled by observing some designed adversarial examples,which are generated by adding perturbations to the original images.In this paper,we propose a novel adversarial example generation method,called DCVAE-adv.Different from the existing methods,DCVAE-adv constructs adversarial examples by mixing both explicit and implicit perturbations without using original images.Furthermore,the proposed method can be applied to both white box and black box attacks.In addition,in the inference stage,the adversarial examples can be generated without loading the original images into memory,which greatly reduces the memory overhead.We compared DCVAE-adv with three most advanced adversarial attack algorithms:FGSM,AdvGAN,and AdvGAN++.The experimental results demonstrate that DCVAE-adv is superior to these state-of-the-art methods in terms of attack success rate and transfer ability for targeted attack.Our code is available at https://github.com/xzforeverlove/DCVAE-adv.

关 键 词：deep neural network adversarial examples white box attack black box attack robustness 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP309[自动化与计算机技术—控制科学与工程]