Using Wasserstein GAN to generate high quality adversarial examples
Xiong, Zhihan
This item is only available for download by members of the University of Illinois community. Students, faculty, and staff at the U of I may log in with your NetID and password to view the item. If you are trying to access an Illinois-restricted dissertation or thesis, you can request a copy through your library's Inter-Library Loan office or purchase a copy directly from ProQuest.
Permalink
https://hdl.handle.net/2142/100047
Description
Title
Using Wasserstein GAN to generate high quality adversarial examples
Author(s)
Xiong, Zhihan
Contributor(s)
Moulin, Pierre
Issue Date
2018-05
Keyword(s)
adversarial machine learning
white-box targeted attack
Wasserstein GAN
neural networks
Abstract
Although Deep Neural Networks (DNNs) have state-of-the-art performance
in various machine learning tasks, in recent years, they are found to be
vulnerable to so-called adversarial examples Specifically, take x is an element of D on
which a neural network has very high classification accuracy. It is possible to
find some small perturbation Δx so that even though the difference between
x and x + Δx = x′ is almost imperceptible to humans, the given neural
network is very likely to incorrectly classify x + Δx.
Several gradient and optimization based methods have been proposed to
create such adversarial examples x′, but many of them cannot achieve high
speed and high quality x′ simultaneously. In this thesis, we propose a new
algorithm to generate adversarial examples based on Generative Adversarial
Networks (GANs), specifically, a modification to the training algorithm of
the Improved Wasserstein GAN. The trained generator is able to create x′
very similar to the original x while keeping the classification accuracy of the
target model as low as the state-of-the-art attack. Furthermore, although
training a GAN might be slow, after it is trained, it can generate adversarial
examples much faster than previous optimization-based methods. Our goal
is for this work to be used for further research on robust neural networks.
Use this login method if you
don't
have an
@illinois.edu
email address.
(Oops, I do have one)
IDEALS migrated to a new platform on June 23, 2022. If you created
your account prior to this date, you will have to reset your password
using the forgot-password link below.