Design and Implementation of Decentralized Data Protection Protocol for Generative AI

Bingxue Zhang, Yang Shi, Feida Zhu, Wang-Chien Lee

Abstract: In recent years, generative AI has made significant advancements. Specifically, in the image domain, one major breakthrough is the adoption of diffusion model-based customized image generation technology, allowing users to generate personalized content based on their own copyrighted images. However, along with this development, the misuse of copyrighted data has become a growing concern. To tackle this issue, this paper first focuses on protecting data from being maliciously trained or inferred by generative AI models in the image domain. Given the limitations of current adversarial example methods, we propose the EMCF (Enhanced Mist, Color space, and Frequency) data protection method with stronger robustness and better visual quality. Furthermore, from the standpoint of AI governance, merely preventing data misuse without considering its circulation and creative reuse demands will fail to unleash its full potential. To bridge this gap, beyond the EMCF data protection method, this paper further incorporates blockchain technology to design a decentralized data protection protocol, ensuring the scalable data circulation within authorized boundaries. Experimental results verify its effectiveness in preventing AI-generated image models from using the protected data for training, providing a new idea for AI governance in the image domain.

External IDs:dblp:conf/dasfaa/ZhangSZL25