Is it fairly easy? Seems useful for a public site like Lemmy and the fediverse
https://nightshade.cs.uchicago.edu/whatis.html
https://decrypt.co/203153/ai-prompt-data-poisoning-nightshared
Is it fairly easy? Seems useful for a public site like Lemmy and the fediverse
https://nightshade.cs.uchicago.edu/whatis.html
https://decrypt.co/203153/ai-prompt-data-poisoning-nightshared
A really fun side effect of stuff like this is when you generate something that looks like a pencil sketch or something, you’ll often get partial pencils in the middle or upper corner of the image because they are quite often photod with pencils on them to indicate the medium.
So even something that simple is sort of poisoning the models. And if they all have that obnoxious signature or QR code, the generators are going to start including those and that’s just gold.
I don’t really think that’s poisoning much. It’s not hard to crop out the pencil after.
It is definitely difficult to get rid of when it’s generated in the middle of intricate detail, which it often is.
I’m not saying it’s the same thing as actually poisoning, but it does negatively impact the resulting generations.
If it’s in the middle of intricate detail it will make it harder to appreciate that detail as a human.
Anyway, it’s easy to make an AI to remove such things. Just take a million images, add watermarks, and train the AI to produce the original images.