结合最大内接圆的图像对抗样本生成算法  

Image Adversarial Examples Generation Algorithm Combined with Maximum Inscribed Circle

作　　者：冯博 刘万平[1] 南海 FENG Bo;LIU Wanping;NAN Hai(College of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学计算机科学与工程学院,重庆400054

出　　处：《小型微型计算机系统》2024年第6期1436-1443,共8页Journal of Chinese Computer Systems

基　　金：重庆市自然科学基金项目(cstc2021jcyj-msxmX0594)资助;重庆市教委科学技术研究项目(KJQN201901133)资助。

摘　　要：深度学习算法已经广泛应用于对抗样本领域.针对图像领域模型生成对抗样本,是发掘图像领域模型的弱点并完善对抗样本检测方法的关键.本文提出一种结合最大内接圆的图像对抗样本生成算法,通过作决策边界的最大内接圆计算出最近决策边界与该圆的切点即为对抗样本点,有效提升了生成对抗样本的成功率和欺骗性.实验使用ImageNet和Cifar10数据集对ResNet18,GoogLeNet,VGG16,MobileNetV2模型生成对抗样本.在本文选取的样本中,ImageNet数据集对这4个模型生成的平均对抗扰动量分别降低了0.1093、0.1697、0.0952、0.0905,Cifar10数据集对这4个模型分别降低了0.0045、0.0049、0.0072、0.0041.这体现了本文方法的优越性与普遍适用性.Deep learning algorithm has been widely used in the field of adversarial examples.Generating adversarial examples for computer vision model is the key way to explore the weaknesses of computer vision model and improve the detection method of adversarial examples.This paper proposes an image adversarial sample generation algorithm combined with maximum inscribed circle,the tangent point between the nearest decision boundary and the circle is the adversarial examples point by drawing the maximum inscribed circle of the decision boundary,which effectively improves the success rate and deception of generating adversarial examples.The experiments use ImageNet and Cifar10 datasets to generate adversarial examples for ResNet18,GoogLeNet,VGG16,MobileNetV2 models.In this paper,the average perturbation generated by ImageNet dataset for these models is reduced by 0.1093,0.1697,0.0952,0.0905 respectively,and Cifar10 dataset is reduced by 0.0045,0.0049,0.0072,0.0041 respectively.This reflects the superiority and universal applicability of the method in this paper.

关 键 词：黑盒攻击 最大内接圆 黑盒决策边界攻击 对抗扰动量 对抗样本 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]