Wasserstein Adversarial Examples via Projected Sinkhorn Iterations

Eric Wong, Frank Schmidt, Zico Kolter
Proceedings of the 36th International Conference on Machine Learning, PMLR 97:6808-6817, 2019.

Abstract

A rapidly growing area of work has studied the existence of adversarial examples, datapoints which have been perturbed to fool a classifier, but the vast majority of these works have focused primarily on threat models defined by $\ell_p$ norm-bounded perturbations. In this paper, we propose a new threat model for adversarial attacks based on the Wasserstein distance. In the image classification setting, such distances measure the cost of moving pixel mass, which can naturally represent “standard” image manipulations such as scaling, rotation, translation, and distortion (and can potentially be applied to other settings as well). To generate Wasserstein adversarial examples, we develop a procedure for approximate projection onto the Wasserstein ball, based upon a modified version of the Sinkhorn iteration. The resulting algorithm can successfully attack image classification models, bringing traditional CIFAR10 models down to 3% accuracy within a Wasserstein ball with radius 0.1 (i.e., moving 10% of the image mass 1 pixel), and we demonstrate that PGD-based adversarial training can improve this adversarial accuracy to 76%. In total, this work opens up a new direction of study in adversarial robustness, more formally considering convex metrics that accurately capture the invariances that we typically believe should exist in classifiers, and code for all experiments in the paper is available at https://github.com/locuslab/projected_sinkhorn.

Cite this Paper

BibTeX
@InProceedings{pmlr-v97-wong19a, title = {{W}asserstein Adversarial Examples via Projected {S}inkhorn Iterations}, author = {Wong, Eric and Schmidt, Frank and Kolter, Zico}, booktitle = {Proceedings of the 36th International Conference on Machine Learning}, pages = {6808--6817}, year = {2019}, editor = {Chaudhuri, Kamalika and Salakhutdinov, Ruslan}, volume = {97}, series = {Proceedings of Machine Learning Research}, month = {09--15 Jun}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v97/wong19a/wong19a.pdf}, url = {https://proceedings.mlr.press/v97/wong19a.html}, abstract = {A rapidly growing area of work has studied the existence of adversarial examples, datapoints which have been perturbed to fool a classifier, but the vast majority of these works have focused primarily on threat models defined by $\ell_p$ norm-bounded perturbations. In this paper, we propose a new threat model for adversarial attacks based on the Wasserstein distance. In the image classification setting, such distances measure the cost of moving pixel mass, which can naturally represent “standard” image manipulations such as scaling, rotation, translation, and distortion (and can potentially be applied to other settings as well). To generate Wasserstein adversarial examples, we develop a procedure for approximate projection onto the Wasserstein ball, based upon a modified version of the Sinkhorn iteration. The resulting algorithm can successfully attack image classification models, bringing traditional CIFAR10 models down to 3% accuracy within a Wasserstein ball with radius 0.1 (i.e., moving 10% of the image mass 1 pixel), and we demonstrate that PGD-based adversarial training can improve this adversarial accuracy to 76%. In total, this work opens up a new direction of study in adversarial robustness, more formally considering convex metrics that accurately capture the invariances that we typically believe should exist in classifiers, and code for all experiments in the paper is available at https://github.com/locuslab/projected_sinkhorn.} }
Endnote
%0 Conference Paper %T Wasserstein Adversarial Examples via Projected Sinkhorn Iterations %A Eric Wong %A Frank Schmidt %A Zico Kolter %B Proceedings of the 36th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2019 %E Kamalika Chaudhuri %E Ruslan Salakhutdinov %F pmlr-v97-wong19a %I PMLR %P 6808--6817 %U https://proceedings.mlr.press/v97/wong19a.html %V 97 %X A rapidly growing area of work has studied the existence of adversarial examples, datapoints which have been perturbed to fool a classifier, but the vast majority of these works have focused primarily on threat models defined by $\ell_p$ norm-bounded perturbations. In this paper, we propose a new threat model for adversarial attacks based on the Wasserstein distance. In the image classification setting, such distances measure the cost of moving pixel mass, which can naturally represent “standard” image manipulations such as scaling, rotation, translation, and distortion (and can potentially be applied to other settings as well). To generate Wasserstein adversarial examples, we develop a procedure for approximate projection onto the Wasserstein ball, based upon a modified version of the Sinkhorn iteration. The resulting algorithm can successfully attack image classification models, bringing traditional CIFAR10 models down to 3% accuracy within a Wasserstein ball with radius 0.1 (i.e., moving 10% of the image mass 1 pixel), and we demonstrate that PGD-based adversarial training can improve this adversarial accuracy to 76%. In total, this work opens up a new direction of study in adversarial robustness, more formally considering convex metrics that accurately capture the invariances that we typically believe should exist in classifiers, and code for all experiments in the paper is available at https://github.com/locuslab/projected_sinkhorn.
APA
Wong, E., Schmidt, F. & Kolter, Z.. (2019). Wasserstein Adversarial Examples via Projected Sinkhorn Iterations. Proceedings of the 36th International Conference on Machine Learning, in Proceedings of Machine Learning Research 97:6808-6817 Available from https://proceedings.mlr.press/v97/wong19a.html.

Related Material