Attribute-Guided Adversarial Training for Robustness to Natural Perturbations
Abstract
We report while existing work in robust deep learning has focused on small pixel-level norm-based perturbations, this may not account for perturbations encountered in several real world settings. In many such cases although test data might not be available, broad specifications about the types of perturbations (such as an unknown degree of rotation) may be known. We consider a setup where robustness is expected over an unseen test domain that is not i.i.d. but deviates from the training domain. While this deviation may not be exactly known, its broad characterization is specified a priori, in terms of attributes. We propose an adversarial training approach which learns to generate new samples so as to maximize exposure of the classifier to the attributes-space, without having access to the data from the test domain. Our adversarial training solves a min-max optimization problem, with the inner maximization generating adversarial perturbations, and the outer minimization finding model parameters by optimizing the loss on adversarial perturbations generated from the inner maximization. We demonstrate the applicability of our approach on three types of naturally occurring perturbations --- object-related shifts, geometric transformations, and common image corruptions. Our approach enables deep neural networks to be robust against a widemore »
- Authors:
-
- Arizona State University, Tempe, AZ (United States)
- Lawrence Livermore National Lab. (LLNL), Livermore, CA (United States)
- Publication Date:
- Research Org.:
- Lawrence Livermore National Lab. (LLNL), Livermore, CA (United States)
- Sponsoring Org.:
- USDOE National Nuclear Security Administration (NNSA); USDOE Laboratory Directed Research and Development (LDRD) Program
- OSTI Identifier:
- 1888097
- Report Number(s):
- LLNL-JRNL-814425
Journal ID: ISSN 2159-5399; 1023019
- Grant/Contract Number:
- AC52-07NA27344
- Resource Type:
- Accepted Manuscript
- Journal Name:
- Proceedings of the AAAI Conference on Artificial Intelligence
- Additional Journal Information:
- Journal Volume: 35; Journal Issue: 9; Conference: 35. AAAI Conference on Artificial Intelligence, Held Virtually, 2-9 Feb 2021; Journal ID: ISSN 2159-5399
- Publisher:
- Association for the Advancement of Artificial Intelligence
- Country of Publication:
- United States
- Language:
- English
- Subject:
- 97 MATHEMATICS AND COMPUTING; advanced learning; robustness
Citation Formats
Gokhale, Tejas, Anirudh, Rushil, Kailkhura, Bhavya, J. Thiagarajan, Jayaraman, Baral, Chitta, and Yang, Yezhou. Attribute-Guided Adversarial Training for Robustness to Natural Perturbations. United States: N. p., 2021.
Web. doi:10.1609/aaai.v35i9.16927.
Gokhale, Tejas, Anirudh, Rushil, Kailkhura, Bhavya, J. Thiagarajan, Jayaraman, Baral, Chitta, & Yang, Yezhou. Attribute-Guided Adversarial Training for Robustness to Natural Perturbations. United States. https://doi.org/10.1609/aaai.v35i9.16927
Gokhale, Tejas, Anirudh, Rushil, Kailkhura, Bhavya, J. Thiagarajan, Jayaraman, Baral, Chitta, and Yang, Yezhou. Tue .
"Attribute-Guided Adversarial Training for Robustness to Natural Perturbations". United States. https://doi.org/10.1609/aaai.v35i9.16927. https://www.osti.gov/servlets/purl/1888097.
@article{osti_1888097,
title = {Attribute-Guided Adversarial Training for Robustness to Natural Perturbations},
author = {Gokhale, Tejas and Anirudh, Rushil and Kailkhura, Bhavya and J. Thiagarajan, Jayaraman and Baral, Chitta and Yang, Yezhou},
abstractNote = {We report while existing work in robust deep learning has focused on small pixel-level norm-based perturbations, this may not account for perturbations encountered in several real world settings. In many such cases although test data might not be available, broad specifications about the types of perturbations (such as an unknown degree of rotation) may be known. We consider a setup where robustness is expected over an unseen test domain that is not i.i.d. but deviates from the training domain. While this deviation may not be exactly known, its broad characterization is specified a priori, in terms of attributes. We propose an adversarial training approach which learns to generate new samples so as to maximize exposure of the classifier to the attributes-space, without having access to the data from the test domain. Our adversarial training solves a min-max optimization problem, with the inner maximization generating adversarial perturbations, and the outer minimization finding model parameters by optimizing the loss on adversarial perturbations generated from the inner maximization. We demonstrate the applicability of our approach on three types of naturally occurring perturbations --- object-related shifts, geometric transformations, and common image corruptions. Our approach enables deep neural networks to be robust against a wide range of naturally occurring perturbations. We demonstrate the usefulness of the proposed approach by showing the robustness gains of deep neural networks trained using our adversarial training on MNIST, CIFAR-10, and a new variant of the CLEVR dataset.},
doi = {10.1609/aaai.v35i9.16927},
journal = {Proceedings of the AAAI Conference on Artificial Intelligence},
number = 9,
volume = 35,
place = {United States},
year = {Tue May 18 00:00:00 EDT 2021},
month = {Tue May 18 00:00:00 EDT 2021}
}