Adversarial attacks in diffusers

[Uzay-G]

I don't have much experience at all with this library and might be misguided with my question, in which case I'd appreciate just getting forwarded to the right links.

If i have a diffusers model loaded, eg a Kandinsky model, I can train or do a zero shot classifier, eg on Imagenet, but is there any existing packaging to then make an adversarial attack on this classifier? eg given an input image perturb it to get the wrong classification label?

If not, how would I start thinking about doing this? What is the right layer of abstraction to do this kind of thing.