#ModelVulnerability

[osintelligence]

https://bit.ly/3saXhuD - 🖼 A new tool, Nightshade, allows artists to alter pixels within their art. If this art is incorporated into an AI training set, the AI model may malfunction in unpredictable ways. The motive is to deter AI companies from using artworks without artists' permissions. The outcome could be, for instance, image-generating AI models producing erroneous outputs, such as turning dogs into cats. #Nightshade #AI #ArtistsRights 📢 Several AI firms like OpenAI, Meta, and Google face legal challenges from artists claiming their works were used without consent. Nightshade, developed under the guidance of Professor Ben Zhao, is seen as a means to give power back to artists by serving as a deterrent against copyright infringement. Neither Meta, Google, Stability AI, nor OpenAI commented on their potential reactions to this tool. #OpenAI #Meta #Google #Copyright 🔒 Alongside Nightshade, Zhao's team also created Glaze. This allows artists to mask their unique style to avoid detection by AI scraping tools. Soon, Nightshade will be integrated into Glaze, and it's set to become open source. The more artists use these tools, the stronger they become, potentially damaging large AI models. #Glaze #OpenSource #AISecurity 🎯 Focusing on the mechanism, Nightshade exploits a weakness in generative AI models which train on vast data sets. By poisoning art data, for instance, an AI might confuse hats with cakes. Removing these corrupted data points is challenging. Experiments showed that with a few hundred poisoned images, AI outputs become significantly distorted. #DataPoisoning #AIModel 🔗 The influence of Nightshade isn't limited to direct keywords. If the AI is trained on a corrupted image labeled as “fantasy art,” related terms such as “dragon” or “Lord of the Rings castle” could also generate skewed outputs. #AIInfluence #KeywordAssociation ⚠️ While Nightshade could be a potent tool for artists, Zhao acknowledges potential misuse. Corrupting larger AI models would require thousands of tainted samples. Experts like Professor Vitaly Shmatikov emphasize that defenses against such attacks are vital. Gautam Kamath praised the research, noting it highlights vulnerabilities in AI models. #AIAbuse #ModelVulnerability 🤝 Nightshade may push AI companies to recognize artists' rights better, potentially leading to increased royalty payments. While some AI firms allow artists to opt out of training models, artists argue this isn't sufficient. Tools like Nightshade and Glaze might restore some power to artists, giving them confidence to showcase their art online without fear.