基于纹理和颜色感知距离的对抗样本生成算法

Adversarial Examples Generation Method Based on Texture and Perceptual Color Distance

理想的对抗样本不仅要成功欺骗机器学习分类器,同时还应不易被人类视觉感知到差异。传统的算法仅采用L_(p)范数衡量对抗样本扰动的大小,往往导致视距差异与感官不匹配等问题。该文提出了一种基于纹理和颜色感知距离的对抗样本生成算法(Aho-λ),其基本原理是尽可能地将扰动嵌入原始图像的高纹理区域,且基于颜色感知距离构建损失函数,从而降低原始图像和对抗样本之间的视距差异,最后利用自适应参数调节算法加快训练的收敛速度。在相近的L_(p)范数和可迁移性情形下,与DDN和C&W算法相比,该算法生成的对抗样本颜色感知距离更低,而且能以更少的迭代次数更快地生成对抗样本。

Ideal adversarial examples should not only successfully deceive the machine learning classifier,but also should not easily be perceived by human vision.In the traditional algorithms,only the norm is adopted as a measurement index of the perturbation size of adversarial examples,which usually leads to the difference in the visibility range.In this paper,a method for adversarial examples generation based on the texture and perceptual color distance is developed.The main idea is to embed the perturbation into a high texture area of an image and optimize the perceptual color distance,so as to reduce the difference in the visibility range between the original image and adversarial example.Moreover,an automatic hyperparameter optimization method is employed to accelerate the convergence of backpropagation.Experimental evaluation shows that the proposed algorithm can obtain the smallest L_(2) norm and perceptual color distance than other algorithms.Meanwhile,a smaller number of iterations was required to obtain adversarial examples.