News for hackers. Stuff that matters.
Re:AI doppelgänger experiment – Part 1: The training by jsheard on Sunday May 19, 2024 @03:08PM (#40409048)
There are active countermeasures in the form of Glaze/Nightshade, but I don't know how effective those turned out to be in practice.
Re:AI doppelgänger experiment – Part 1: The training by ccgreg on Sunday May 19, 2024 @04:17PM (#40409465)
I wonder if Glaze/Nightshade makes it difficult for software to describe the image for a blind person?
Re:AI doppelgänger experiment – Part 1: The training by GaggiX on Sunday May 19, 2024 @04:31PM (#40409537)
Not really, because Nightshade should have made image labeling more difficult, but if you try it, you'll see that it doesn't do anything; multimodal models are too powerful nowadays to be fooled by small adversarial noise generated using CLIP LPIPS (small enough not to be too noticeable to us).

And Glaze does not try to interfere with labeling.

Re:AI doppelgänger experiment – Part 1: The training by astrange on Sunday May 19, 2024 @05:14PM (#40409798)
They don't work, and nothing in the category can ever truly work.

Models and adversarial data are equally powerful - you can find an adversarial example for any current model, but you can also train a model that can handle any existing adversarial data.

Any image is a good example of /something/ - at worst it's only an example of itself and irrelevant for any other class. Don't know if there's a name for this principle but it's kind of like a Church-Turing thesis for data.