AI Image Grabber: A Tool for Image Extraction

Researchers at the University of Chicago have announced the development of a tool called "Nightshade" that can "poison" the graphics used by artificial intelligence companies to train image generation models. Nightshade manipulates image pixels to alter the output during training. These changes are imperceptible to the human eye before processing.

One of the authors of the paper "Specific Prompt Poisoning Attacks on Text-to-Image Generation Models," Ben Zhao, states that Nightshade can manipulate the data to transform images of dogs into cats during training, among other transformations such as cars into cars and hats into cakes. This work has been published on the arXiv preprint server.

Ben Zhao believes that a moderate number of Nightshade attacks can disrupt the general features of text-to-image generation models, effectively rendering them unable to generate meaningful images.

He refers to his team's creation as the "last line of defense against content creators who ignore crawling instructions/crawling directives from web crawlers."

For a long time, artists have been concerned that companies like Google, OpenAI, Stable AI, and Meta have been collecting billions of online images for training image generation tools without compensating the creators.

Eva Toorenent, an advisor to the European Guild of Artificial Intelligence Regulation in the Netherlands, said this practice "extracts the creativity of millions of artists."

"It's absolutely terrifying," she said in a recent interview.

Ben Zhao's team has demonstrated that despite the common belief that disrupting crawling operations requires uploading a large number of modified images, they can achieve disruption using fewer than 100 "poison" samples. They achieve this through specific prompt poisoning attacks, which require far fewer samples than the model training dataset.

Ben Zhao believes Nightshade is not only useful for individual artists but also for large companies such as film studios and game developers.

"For example, Disney could apply Nightshade to their print images of 'Cinderella' while coordinating the poisoning concept of 'mermaids' with others," said Ben Zhao.

Nightshade can also alter artistic styles. For example, prompts requesting creations in the Baroque style may produce images in the Cubist style.

This tool emerges amidst opposition to AI companies encroaching on online content. Based on fair use rules claimed by these companies, people filed lawsuits against Google and Microsoft's OpenAI last year, accusing these tech giants of using copyrighted materials without authorization to train their AI systems.

Plaintiff lawyer Ryan Clarkson stated, "Google does not own the internet, nor does it own our creative works, our pictures of ourselves, our families, and our children, or anything else just because we share it online." If found guilty, these companies could face billions of dollars in fines.

Google is seeking to dismiss the lawsuit, stating in court documents, "Learning from publicly available information is not theft, invasion of privacy, conversion, negligence, unfair competition, or copyright infringement."

According to Toorenent, Nightshade "will make (AI companies) think twice because they risk taking away our works without our consent, thereby disrupting their entire model."