Abstract:
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
L2 norm and perceptual color distance than other algorithms. Meanwhile, a smaller number of iterations was required to obtain adversarial examples