Implement low frequency adversarial perturbation strategy

[ririnicolae]

This is a black-box attack search strategy allowing to reduce the number of queries to a model by 2-4x. It is based on discrete cosine transform (DCT). This selection strategy can be combined with existing black-box attacks, e.g., the paper proves effectiveness with NES and boundary attack. This should be implemented as a wrapper if possible, to allow the usage with existing black-box attacks in ART.

Paper link: https://arxiv.org/abs/1809.08758

[rish-16]

@beat-buesser Hello there, could you please assign this issue to me? I'll get started on it.

[beat-buesser]

@rish-16 Thank you! And let us know here if you have questions or need support.

[rish-16]

Will do. Thank you

[beat-buesser]

Hi @rish-16 Have you been able to make progress with this issue?

[beat-buesser]

Hi @rish-16 Are you still working on this issue?

[kztakemoto]

@ririnicolae Our implementation in ART I forked may be useful.
https://github.com/kztakemoto/SimBA-Keras

[ririnicolae]

@kztakemoto Would you like to pull request to ART your implementation of this attack?

[kztakemoto]

@ririnicolae thanks. I pull requested. #469