Are Perceptually-Aligned Gradients a General Property of Robust
Classifiers?
S. Kaur, J. Cohen, and Z. Lipton. (2019)cite arxiv:1910.08640Comment: To appear in the "Science Meets Engineering of Deep Learning" Workshop at NeurIPS 2019.
Abstract
For a standard convolutional neural network, optimizing over the input pixels
to maximize the score of some target class will generally produce a
grainy-looking version of the original image. However, researchers have
demonstrated that for adversarially-trained neural networks, this optimization
produces images that uncannily resemble the target class. In this paper, we
show that these "perceptually-aligned gradients" also occur under randomized
smoothing, an alternative means of constructing adversarially-robust
classifiers. Our finding suggests that perceptually-aligned gradients may be a
general property of robust classifiers, rather than a specific property of
adversarially-trained neural networks. We hope that our results will inspire
research aimed at explaining this link between perceptually-aligned gradients
and adversarial robustness.
Description
[1910.08640] Are Perceptually-Aligned Gradients a General Property of Robust Classifiers?
%0 Journal Article
%1 kaur2019perceptuallyaligned
%A Kaur, Simran
%A Cohen, Jeremy
%A Lipton, Zachary C.
%D 2019
%K robustness
%T Are Perceptually-Aligned Gradients a General Property of Robust
Classifiers?
%U http://arxiv.org/abs/1910.08640
%X For a standard convolutional neural network, optimizing over the input pixels
to maximize the score of some target class will generally produce a
grainy-looking version of the original image. However, researchers have
demonstrated that for adversarially-trained neural networks, this optimization
produces images that uncannily resemble the target class. In this paper, we
show that these "perceptually-aligned gradients" also occur under randomized
smoothing, an alternative means of constructing adversarially-robust
classifiers. Our finding suggests that perceptually-aligned gradients may be a
general property of robust classifiers, rather than a specific property of
adversarially-trained neural networks. We hope that our results will inspire
research aimed at explaining this link between perceptually-aligned gradients
and adversarial robustness.
@article{kaur2019perceptuallyaligned,
abstract = {For a standard convolutional neural network, optimizing over the input pixels
to maximize the score of some target class will generally produce a
grainy-looking version of the original image. However, researchers have
demonstrated that for adversarially-trained neural networks, this optimization
produces images that uncannily resemble the target class. In this paper, we
show that these "perceptually-aligned gradients" also occur under randomized
smoothing, an alternative means of constructing adversarially-robust
classifiers. Our finding suggests that perceptually-aligned gradients may be a
general property of robust classifiers, rather than a specific property of
adversarially-trained neural networks. We hope that our results will inspire
research aimed at explaining this link between perceptually-aligned gradients
and adversarial robustness.},
added-at = {2019-10-22T20:45:03.000+0200},
author = {Kaur, Simran and Cohen, Jeremy and Lipton, Zachary C.},
biburl = {https://www.bibsonomy.org/bibtex/2f06143b1a97c0a994a0974474555fc33/kirk86},
description = {[1910.08640] Are Perceptually-Aligned Gradients a General Property of Robust Classifiers?},
interhash = {b6dea3f2615543d8cca49e91cb14149f},
intrahash = {f06143b1a97c0a994a0974474555fc33},
keywords = {robustness},
note = {cite arxiv:1910.08640Comment: To appear in the "Science Meets Engineering of Deep Learning" Workshop at NeurIPS 2019},
timestamp = {2019-10-22T20:45:03.000+0200},
title = {Are Perceptually-Aligned Gradients a General Property of Robust
Classifiers?},
url = {http://arxiv.org/abs/1910.08640},
year = 2019
}