Robust watermarking for diffusion model generated images

Abstract

With the wide application of diffusion models in the field of image generation, image copyright protection and traceability have become increasingly complex and challenging. To address these problems, this paper proposes a robust watermarking method for diffusion model generated images to achieve their copyright protection and traceability. The method designs an invertible mapping module to replicate and cryptographically map the watermark information into an approximately Gaussian distributed noise, which is highly consistent with the distribution of the original generation model. The mapped watermark noise serves as the latent space vector of the generative model, preserving both image generation quality and model performance. In the watermark extraction stage, the original watermark information can be accurately recovered from the generated image through the reverse extraction and voting mechanism. Experimental results show that the proposed method demonstrates excellent performance in terms of image watermark extraction accuracy, robustness and watermark image generation quality. It can still maintain 99 % true positive rate and 97.5 % bit accuracy under various attacks, and the overall performance in the detection and traceability scenarios is significantly better than the existing baseline methods.