0%

(ICLR 2018) Defense-GAN:Protecting classifiers against adversarial attacks using generative models

Samangouei P, Kabkab M, Chellappa R. Defense-gan: Protecting classifiers against adversarial attacks using generative models[J]. arXiv preprint arXiv:1805.06605, 2018.



1. Overview


In this paper, it proposed Defense-GAN methods

  • train to model the distribution of unperturbed images
  • does not assume knowledge of the process for generating the adversarial examples
  • effective against both white-box and black-box attacks

1.1. Defense Type

  • modify the training data: adversarial training
    modify the training procedure of the classifier to reduce the magnitude of gradients: defensive distilation
    remove the adversarial noise

1.2. Attack Type

  • FGSM
  • RAND+FGSM



  • C&W



  • Iterative FGSM
  • Jacobian-based Saliency Map Attack (JSMA)
  • Deepfool

1.3. Algorithm