We have talked a lot about how AI is grabbing art from all over the internet to create its own creations. Obviously, these are not original, just a mishmash of existing images. This drives millions of artists crazy, who could now take revenge in the best way possible.
Artists could soon have a new weapon to prevent their works from becoming cannon fodder for machine learning.
The tool, called Nightshade, introduces small changes to the pixels of a digital artwork to “poison” it and make it useless for AI training.
This is how a tool that could change the current paradigm works
The MIT Technology Review reports that a team led by Ben Zhao, a professor at the University of Chicago, presented Nightshade for review at the USENIX computer security conference.
The program works by making small modifications to an image that, although invisible to the human eye, cause AI algorithms to identify them completely incorrectly.
For example, an artist can paint a picture of a cat that any human or AI examining it can clearly identify as a feline. However, by applying Nightshade, humans will still see the same image while the AI will mistakenly believe it is a dog.
And they explain that if AI is inundated with enough bad training material like this, soon a request for an image of a cat will make it generate a dog instead.
Of course, a single poisoned image is unlikely to have a significant effect on an AI image generator’s algorithm. Its training data would need to be contaminated by thousands of altered images before a real impact could be observed.
However, it is known that AI image generators indiscriminately collect thousands of new samples from the internet to refine their algorithm. If a sufficient number of artists upload their images with Nightshade applied, they could potentially render these AI tools useless.
