Comment by gweinberg
3 years ago
For this to work, wouldn't you have to have an enormous number of artists collaborating on "poisoning" their images the same way (cow to handbag) while somehow keeping it secret form ai trainers that they were doing this? It seems to me that even if the technology works perfectly as intended, you're effectively just mislabeling a tiny fraction of the training data.
1. They don't need an enormous number of artists; the research paper showed significant results with even 50 poisoned image samples in the dataset, which is enough to be contained in even a single artist's online gallery.
2. They don't need to keep it a secret; the goal is to remove these images from the training data, in a way that would be much more efficient than simply adding a "please don't include my art in your ai scraper" message next to your pictures.