eduzhai > Applied Sciences > Engineering >

Color and Edge-Aware Adversarial Image Perturbations

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 11 pages

Abstract: Adversarial perturbation of images, in which a source image is deliberatelymodified with the intent of causing a classifier to misclassify the image,provides important insight into the robustness of image classifiers. In thiswork we develop two new methods for constructing adversarial perturbations,both of which are motivated by minimizing human ability to detect changesbetween the perturbed and source image. The first of these, the Edge-Awaremethod, reduces the magnitude of perturbations permitted in smooth regions ofan image where changes are more easily detected. Our second method, theColor-Aware method, performs the perturbation in a color space which accuratelycaptures human ability to distinguish differences in colors, thus reducing theperceived change. The Color-Aware and Edge-Aware methods can also beimplemented simultaneously, resulting in image perturbations which account forboth human color perception and sensitivity to changes in homogeneous regions.Because Edge-Aware and Color-Aware modifications exist for many imageperturbations techniques, we also focus on computation to demonstrate theirpotential for use within more complex perturbation schemes. We empiricallydemonstrate that the Color-Aware and Edge-Aware perturbations we considereffectively cause misclassification, are less distinguishable to humanperception, and are as easy to compute as the most efficient image perturbationtechniques. Code and demo available atthis https URL

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×