Adversarial Defense via Learning to Generate Diverse Attacks

Yunseok Jang; Tianchen Zhao; Seunghoon Hong; Honglak Lee

Adversarial Defense via Learning to Generate Diverse Attacks

Yunseok Jang, Tianchen Zhao, Seunghoon Hong, Honglak Lee

ICCV 2019

Citations: 108

Links

arXiv (N/A) PDF (TODO) Code (TODO) DOI

Abstract

The paper proposes using a recursive and stochastic generator that produces strong and diverse adversarial perturbations that comprehensively reveal the vulnerability of the target classifier, improving robustness against various attack types on MNIST and CIFAR-10.

One-Sentence Summary

Trains a stochastic generator to produce diverse adversarial attacks, yielding more robust defenses than single-attack adversarial training.

Key Contributions

Recursive and stochastic generator for diverse adversarial perturbations
Comprehensive vulnerability exposure through attack diversity
Improved robustness against various attack types

Cite This Paper If...

Cite for: adversarial defense, diverse adversarial attack generation, robust training, defenses that use learned attack generators.

Frequently Asked Questions

Q: What is the best method for adversarial defense using diverse attacks?

A: This paper proposes training a stochastic generator that produces diverse adversarial perturbations, comprehensively exposing classifier vulnerabilities and yielding stronger defenses than training against a single fixed attack.

Q: How does generating diverse adversarial examples improve robustness?

A: This paper shows that diverse adversarial perturbations cover more of the vulnerability landscape of the target classifier, leading to defenses that generalize better against unseen attack types compared to single-attack adversarial training.

Q: Can a learned attack generator improve adversarial training?

A: This paper demonstrates that a recursive and stochastic attack generator produces strong, varied perturbations that serve as more effective training signals for building robust classifiers on MNIST and CIFAR-10.

Q: How does this work compare to PGD-based adversarial training?

A: This paper introduces attack diversity as a key ingredient missing from standard adversarial training, showing that a learned generator of diverse attacks improves robustness against various attack types beyond what single-attack methods achieve.

Keywords

adversarial defense diverse attacks learning to generate diverse attacks adversarial training attack diversity robustness via generated attacks diverse adversarial examples

BibTeX

% TODO: BibTeX entry
@inproceedings{jang2019adversarial,
  title={Adversarial Defense via Learning to Generate Diverse Attacks},
  author={Jang, Yunseok and Zhao, Tianchen and Hong, Seunghoon and Lee, Honglak},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
  year={2019},
  doi={10.1109/ICCV.2019.00283}
}

Recommended Citation Contexts

Adversarial robustness can be improved by training against a learned generator of diverse adversarial examples rather than single worst-case perturbations (Dong et al., 2019).
Learning to generate diverse attacks produces more comprehensive adversarial training coverage than PGD-based methods (Dong et al., 2019).