University of Illinois Urbana-Champaign

Certified adversarial robustness via randomized discretization

Vijitbenjaronk, W. Duke

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/2142/104039

Description

Title
Certified adversarial robustness via randomized discretization
Author(s)
Vijitbenjaronk, W. Duke
Contributor(s)
Telgarsky, Matus
Issue Date
2019-05
Keyword(s)
  • adversarial robustness
  • randomized discretion
  • accuracy of deep learning models
Abstract
Modern machine learning algorithms are able to reach an astonishingly high level of performance in a variety of useful tasks. However, small adversarial perturbations have been shown to drastically reduce the accuracy of deep learning models not specifically trained to resist them. This problem is of practical significance due to security concerns about models deployed in industry, and of theoretical significance due to the connections between this problem and the underlying themes of optimization and generalization. In this paper, we propose and analyze a simple and computationally efficient defense against adversarial attacks based on randomized discretization to a relatively small set of points that is agnostic of the underlying classifier. We show that that this strategy leads to a lower bound on the classification accuracy using tools from computational geometry and information theory. Unlike prior work, the proposed strategy allows for easily estimable data-dependent accuracy guarantees at inference time, and demonstrates a weaker dependence on the dimensionality of its inputs.
Type of Resource
text
Language
en
Permalink
http://hdl.handle.net/2142/104039

Owning Collections

Senior Theses - Electrical and Computer Engineering PRIMARY
The best of ECE undergraduate research