Computer researchers from the University of Chicago have introduced a novel tool named Nightshade designed to “poison” digital artwork. This tool aims to impede the training of image-generating AI models involved in intellectual property theft, such as DALL-E, Midjourney, and Stable Diffusion.
Origins of Nightshade: Nightshade emerged from the Glaze Project, led by a team of computer scientists at the University of Chicago under the guidance of professor Ben Zhao. Previously, the group created Glaze, a tool that distorts how AI training algorithms perceive the style of digital artwork to bewilder the models.
Functionality: Nightshade utilizes the Pytorch open-source machine learning framework to label images at the pixel level. These labels are imperceptible to humans but affect AI models, disrupting the images’ utility for training purposes.
Distinguishing Glaze from Nightshade: While Glaze alters the perceived artistic style of images to confuse AI models, Nightshade manipulates the content of images to mislead the models. For instance, Nightshade could make an AI model interpret a photo of a cat as a photo of a dog. This misleads the model during training, resulting in unexpected outputs.
Challenges: The processes of glazing and shading introduce noise to digital images, with varying levels of distortion that can be adjusted by the user.
Future of Image-Generating Software: Despite these developments, the Glaze Project is not opposed to AI. Instead, it aims to establish a system where users of image-generating programs must obtain approval from rightsholders to access unaltered training images.
Motivation behind the Glaze Project: Professor Zhao emphasizes that the primary objectives of the project are research discovery and positive societal impact, rather than financial gain. The team’s ethos focuses on contributing to knowledge and societal welfare without pursuing profit or commercial ventures.
