基于Targeted-UAP算法的SAR图像对抗样本生成方法

SAR image adversarial sample generation method based on Targeted-UAP algorithm

深度卷积神经网络被广泛应用于合成孔径雷达(SAR)图像目标识别。但有关研究表明,其易受SAR图像上轻微扰动的攻击,从而导致检测、分类识别的失败。为此,首先分析了光学图像下通用对抗扰动(UAP)算法,通过修改其迭代约束条件,提出有目标的UAP(Targeted-UAP)算法,搜索将对抗样本推入目标类别分类边界的最小扰动,生成有目标通用对抗扰动,以实现SAR目标识别网络的有目标攻击。使用Targeted-UAP算法生成有目标攻击SAR图像对抗样本,并在LeNet、VGGNet16、ResNet18三个经典识别模型上验证了其有效性。

Deep convolutional neural networks are widely used in synthetic aperture radar(SAR)image target recognition.However,related studies show that it is vulnerable to be attacted by slight perturbation on SAR images,which leads to the failure of detection,classification and recognition.Based on this,firstly the universal adversarial perturbation(UAP)algorithm is analyzed under the optical image,and the Targeted-UAP algorithm is proposed by modifying its iteration constraints conditions,the smallest perturbation is searched by pushing the adversarial example into the classification boundary of the target category,and a targeted universal adversarial perturbation is generated to achieve targeted attack on the SAR target recognition network.The Targeted-UAP algorithm is used to generate targeted attack adversarial examples of SAR images,its effectiveness is verified on the three classic recognition models of LeNet,VGGNet16,and ResNet18.