ai-pocalypse Computer scientists say they have devised a method to remove image-based protection mechanisms developed to protect artists from unnecessary use of work for AI training.
Some visual artists who are concerned about copyright violations, some visual artists who are concerned about the possibility that AI-generated images could destroy job demand, some visual artists who have been taken to use software that adds “hostile perturbations.” Now, researchers have explained how to beat those perturbations in their paper [PDF] Title: “Lightshed: defeating perturbation-based image copyright protection.”
“We view it primarily as a recognition effort to highlight the weaknesses of existing protection schemes, and we also carried out work under strict responsible disclosure protocols that are required at meetings.” Register.
Lightshed aims to be an antidote to image-based data addiction schemes such as Glaze and Nightshade, developed by University of Chicago computer scientists to block the non-consensual use of artists' works for AI training.
Glaze is software that can be applied to images, making it difficult for machine learning models to mimic the art style of images. Nightshade toxic images with adversarial data (perturbations) that will make AI models misrecognize image functions. Similar proposed defenses against AI data predation include MIST and Metacloak.
Lightshed creators don't try to make it easier for AI companies to bypass defenses against AI training. This is an area of growing industry interest due to the lack of clear legal rules governing AI training (input) and inference (output). Rather, they say that their work aims to demonstrate the inadequacy of existing image protection technologies, which means further improvements can be made.
“Lightshed leverages the wide availability of these protection schemes to generate examples of addiction and model their properties,” the paper explains. “Fingerprints derived from this process allow Lightshed to efficiently extract and neutralize perturbations from protected images.”
By analyzing the toxins, the authors state that their approach allows them to recognize and reverse data addiction patterns. The results are not at all surprising. Other machine learning researchers have discovered that they can remove image watermarks – adding data to track model output rather than messing around with them.
Our method reliably detects and eliminates addiction (perturbation) to the extent that it allows for training on copyrighted images
“Our method reliably detects and eliminates addiction (perturbation) to the extent that it allows training on copyrighted images,” explained Murtuza Jadliwala, an associate professor of computer science at San Antonio, in an email.
“Lightsheds can almost reconstruct trained poisons (perturbations) to demonstrate strong performance, for example, with night shades and glazes, but as mentioned in the paper, the accuracy of reconstructing toxicity that we have never seen before, such as metacloaks.
Other authors of the paper are Sasha Beleuzi, Philip Rieger and Ahmad Reza Sadegi of Darmstadt Institute of Technology. It will be presented at the 34th USENIX Security Symposium in August.
Scrapers and makers
Anti-AI technology represents a response to companies such as AI models such as Midjourney, Openai, and Google Training AI models, which generally sells the ability to mimic that artwork, and sells to those that potentially undermine the income and career of an artist. The visual artist, along with writers, publishers and software developers, sued various AI companies to prevent such use.
