Data poisoning tool lets artists fight back against AI scraping. Here's how

Researchers at the University of Chicago have developed Nightshade, a defensive tool that allows visual artists to protect their work from unauthorized generative AI training. By embedding invisible pixel alterations into digital images, the tool “poisons” the data used to train models. These subtle changes cause significant confusion within the AI, leading it to misidentify subjects and produce chaotic, unpredictable, and useless outputs that fail to match user prompts. The effectiveness of this approach lies in its ability to corrupt specific terms and bleed over into associated concepts and synonyms. Consequently, removing the corrupted data is incredibly difficult for developers, as it requires identifying and deleting individual samples rather than filtering simple noise. This complexity serves as a strong deterrent against indiscriminate data scraping, forcing AI companies to reconsider their reliance on unlicensed creative works and encouraging greater caution when utilizing generative models. This development is highly relevant to open data discussions as it highlights the critical tension between the open availability of digital information for AI training and the rights of individual creators. It underscores the need for transparent, ethical data sourcing practices and demonstrates how technical safeguards can empower individuals to control their digital footprints. As the open data movement evolves, tools like Nightshade illustrate the necessity of balancing innovation with respect for intellectual property, pushing the industry toward more sustainable and consent-based data ecosystems.

Source: zdnet.com
Published on 2023-10-25

AMD adds support for super-popular open source AI tool to its most powerful GPU – now imagine what would happen if an AMD APU could run Stable Diffusion out of the box »

Open Data

Related news