Support for hyperrectangle input for FGSM and PGD

jaypmorgan / Adversarial.jl

Adversarial attacks for Neural Networks written with FluxML

https://jaypmorgan.github.io/Adversarial.jl/dev/

15 stars 1 forks source link

Support for hyperrectangle input for FGSM and PGD #1

Open castrong opened 4 years ago

castrong commented 4 years ago

It seems like FGSM and PGD assume a hypercube input (from the uniform radius parameter ϵ). These algorithms both seem like they could support a hyper-rectangle input with small modifications. I was wondering whether this is true or if I'm missing something about the algorithms. If it is, I could work on a PR to add this functionality.

jaypmorgan commented 4 years ago

It is true that the current implementations use a uniform radius. If you would like to create a function (with the same name) that supports non-uniform radii as you suggest, I would be happy to review the pull request!