Nightshade – With this new tool, artists are now fighting back against AI companies

Artificial intelligence is increasingly being used in the art world, which on the one hand opens up new creative possibilities, but on the other hand displeases many traditional artists.

Generative AI and AI tools have added a completely new dimension to art . They can imitate the style of famous artists, creating new works or even original, unique pieces that can win awards.

But not everyone likes this approach.

With Nightshade and poisoned images, artists are now fighting back against AI companies and Generative AI

Researchers at the University of Chicago have Nightshade, a new tool that allows artists to defend AI models reported MIT Technology Review

Nightshade subtly alters training data for image-generating AI models, thereby launching targeted attacks on these technologies.

This innovative tool allows artists to make invisible changes to the pixels of their art before publishing it online. However, this carries the risk that when incorporated into an AI training set, the resulting model will be corrupted in unpredictable and chaotic ways.

The deliberate “poisoning” of this training data could seriously harm future versions of image-generating AI models such as DALL-E , Midjourney , and Stable Diffusion . Some results could become unusable – dogs could turn into cats, cars into cows, and so on.

MIT Technology Review had exclusive access to a research study that was submitted for peer review at the Usenix

AI companies like OpenAI, Meta, Google , and Stability AI are facing a number of lawsuits from artists who claim that their copyrighted material and personal data were used without consent or compensation.

Ben Zhao , a professor at the University of Chicago and head of the team behind Nightshade, hopes to restore the balance of power between AI companies and artists .

This is to be achieved through a strong deterrent against the infringement of artists' copyrights and intellectual property rights. The companies Meta, Google, Stability AI, and OpenAI have not yet responded to MIT Technology Review's request for comment regarding their potential response.

Nightshade and Glaze as a combined weapon against copyright infringements by large AI companies

Zhao's team also Glaze – a tool for artists – that allows them to "mask" , preventing it from being copied by AI companies.

Similar to Nightshade, subtle changes are made to the pixels of an image. These are invisible to the human eye; however, they manipulate machine learning models to such an extent that they interpret the image as something different than it actually is.

The team plans to integrate the data tool Nightshade into Glaze . Artists will have the option to choose whether or not to use this tool. Additionally, Nightshade will also be offered as open source, allowing others to tinker with it and create their own versions.

By using and creating different versions, the tool becomes increasingly powerful, Zhao emphasizes. The datasets for large AI models can consist of billions of images. The more toxic images that can be fed into the model, the greater the damage caused by this technique.

How do Nightshade and Glaze work?

Artists who want to share their work online but don't want AI companies to use it for their purposes have a solution called Glaze ( click here for the tool ).

Here, users can upload their images and distort them with a different style. Additionally, it is now also possible to use Nightshade

When AI developers scour the internet to gather data for improving or creating models, they are given these manipulated samples, leading to malfunctions.

These contaminated data samples have the effect of influencing AI models in such a way that, for example, hats are recognized as cakes and handbags as toasters. It is extremely difficult to remove such false information – technology companies would have to painstakingly find and delete each individual corrupted sample.

The researchers tested this attack on both the latest models of Stable Diffusion and a self-developed AI model. When they fed Stable Diffusion just 50 poisoned dog images and then asked it to generate more dog images, the result looked strange: creatures with an excessive number of limbs and cartoonish faces emerged.

With 300 poisoned samples, an attacker can even make Stable Diffusion create cat pictures instead of dog pictures.

What impact do these tools have on the power dynamic between creators and AI companies?

Junfeng Yang , a computer science professor at Columbia University who has researched the security of deep learning systems and was not involved in the project, convincingly explains that Nightshade could potentially have a major impact – particularly as a deterrent.

For example, AI companies be encouraged to better respect the rights of artists and be more willing to pay licensing fees .

AI companies like Stability AI and OpenAI, which text-to-image generating models, already offer artists the option to opt out of having their images used to train future versions of these models. However, many artists believe this is insufficient.

Eva Toorenent is an illustrator and artist working under the pseudonym Evaboneva who has used Glaze and critically notes that such opt-out methods would still require submission to technology companies.

Toorenent hopes that Nightshade can change the status quo:

It will make [AI companies] rethink their approach, as they are now able to destroy their entire model by using our work without consent.”

she commented to the MIT Technology Review.

Another artist, Autumn Beverly, reported that tools like Nightshade and Glaze restored her confidence to republish her work online. She had previously removed it after discovering it had been copied LAION

I am truly grateful to have a tool available that can help artists regain control over their work.”

she says.

Joachim Rodriguez y Romero

Owner and Managing Director of Kunstplaza . Publicist, editor, and passionate blogger in the fields of art, design, and creativity since 2011. Graduated with a degree in web design from university (2008). Further developed creative techniques through courses in freehand drawing, expressive painting, and theatre/acting. Profound knowledge of the art market gained through years of journalistic research and numerous collaborations with key players and institutions in the arts and culture sector.

www.kunstplaza.de