New protection from unwanted AI-training

Game changer in protecting images from generative AI training! The tool “poisons” the image, which misleads the AI and can cause malfunction.

➤Background:

👉OpenAI, StabilityAI, Midjourney, Meta, and others train their generative AI models on data they scrape from the internet. That includes content that is copyright protected.

👉Multiple lawsuits are ongoing.

👉Other than lawsuits people couldn’t stop companies from abusing their images.

➤ But now there is a way to protect images!

💣The new tool is called “Nightshade”.

💣Here’s how it works:

Choose an image

Run the image through the tool

The tool modifies pixels in the images

The resulting image looks the same to humans

But, if trained on, it will teach the AI the wrong thing

➤ Examples

AI can be poisoned to give

💣images of dogs when prompted for cats

💣Anime-style painting when prompted for cubism

➤A small number of poisoned images is enough to damage the model

💣You can see examples in the image attached to this post.

💣300 poisoned images were enough to make Stable Diffusion generate images of cats when prompted for a dog.

➤ This work is a game changer!

🔥As far as I know, the tool is still experimental and the research paper is currently under review.

🔥However, the possibility of poisoning generative AI can be revolutionary. This work may inspire many others!

➤ Join the discussion on my LinkedIn post!