Anonymous-Diffusion: Blockchain-Based Privacy-Preserving Stable Diffusion

The field of generative AI is currently seeing a surge in text-to-image generation. Among open-source projects, Stable Diffusion stands out as the state-of-the-art. Artists and service providers often customize diffusion models for unique textures. However, there is a lack of privacy protection for users' input text prompts, output images, and customized models on servers. Ensuring privacy is essential for user trust and safeguarding intellectual property. Current privacy-preserving diffusion models rely on fully homomorphic encryption (FHE), which is time-intensive and can compromise image quality. We introduce Anonymous-Diffusion, a diffusion as a service (DAAS) framework. This framework maintains privacy without using FHE by exploiting the irreversible nature of neural network layers and the characteristic that predicted noise in the diffusion process follows a normalized Gaussian distribution. Furthermore, we ensure anonymity by Smart Contract and Blockchain. User can use this service on demand anonymously. In comparison to existing research like HE-diffusion, which incurs a 200% time overhead and noticeable quality degradation, our protocol achieves the same level of security with only a 4% time overhead and no loss in image quality. To our knowledge, this is the first solution to achieve these results without FHE while preserving high-quality image output.