Skip to content

A repository for work on adversarial image perturbations which respect human perception.

License

Notifications You must be signed in to change notification settings

rbassett3/Color-and-Edge-Aware-Perturbations

Repository files navigation

Color and Edge-Aware Adversarial Perturbations.

This is a repository containing examples and source code for constructing adversarial image perturbations which accurately model human perception of color and texture.

The Edge Aware portion of Color and Edge-Aware Perturbations

The image below compares our method to a multiple iteration version of the Fast Gradient Sign Method (Goodfellow et al., 2014), where our Color-and-Edge-Aware perturbation preserves the smooth texture of the sky while still disrupting the performance of the classifier.

Comparison of FGSM to Color-and-Edge-Aware

If you find our work useful, please cite the following paper, where you can also find more information on the methods we developed.

Robert Bassett, Mitchell Graves, Patrick Reilly. "Color and Edge-Aware Adversarial Image Perturbations." arXiv preprint arXiv:2008.12454 (2020).

Further background information on the topic of adversarial perturbations can be found in the following MS thesis.

Graves, Mitchell. Image Perturbation Generation: Exploring New Ways for Adversaries to Interrupt Neural Network Image Classifiers. NAVAL POSTGRADUATE SCHOOL MONTEREY CA, 2020.

About

A repository for work on adversarial image perturbations which respect human perception.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published