This tool could “fully shield” artists from AI scraping their artwork
While other existing tools allow artists to “poison” a dataset once their image is scraped, Kin.art lets artists completely fly under the radar of AI models.
- Liz Gorny
- 23 January 2024
A new tool could give artists better protection from AI data scraping, the fast-growing problem whereby a model learns how to imitate an “artwork style” from a creative’s existing work without permission. The technology from Kin.art is meant to work proactively, preventing an artwork from ever entering an AI dataset by disrupting the system of labels and images that is needed for an AI to ‘learn’ about an artwork.
While there are existing tools out there fighting data scraping, most of them are based around data poisoning. (Data poisoning is when changes are introduced to an image’s data which damage the way that artwork will be reproduced by AI, eventually causing outputs to mutate). But, data poisoning is a “really expensive computational process requiring expensive graphics cards,” says Flor Ronsmans De Vry, CTO at Kin.art.
“While [tools like] Nightshade and Glaze try to mitigate damage when it’s already too late, we play the prevention card and prevent your art from being inserted successfully in the first place.” The tool is also a more lightweight solution than other data poisoning methods, which means Kin.art can offer it for free.
It works by disrupting how AI understands an image. AI datasets need to pair images with labels to properly train a model – for example, the label ‘dog’ paired with an image of a dog. If these are disrupted, “we can prevent the artwork from being successfully inserted into the dataset”, says Ronsmans De Vry.
GalleryThe image segmentation process (Copyright © Kin.art, 2024)
Kin.art uses two methods to ensure this; label fuzzing, which stops appropriate labels being associated with each image, and something called image segmentation. This is when an image is cut into several irregular pieces before being sent to your browser and reassembled – this throws off AI scrapers who are left with only little pieces of artwork to insert into a dataset. “This dual approach guarantees that artists who showcase their portfolios on Kin.art are fully shielded from unauthorised AI training of their work,” says Ronsmans De Vry.
The tool will be available for free to anyone who uses the portfolio-hosting platform Kin.art from 23 January – if your artwork is uploaded to other places online, it will not be protected from being scraped from those locations.
The protective tool will initially live on Kin.art but the platform has plans to create a service which allows small sites and big platforms to protect themselves from unauthorised use of their media.
Copyright © Kin.art, 2024