基于精确扩散反演的生成式图像内生水印方法  

Generative image endogenous watermarking method based on exact diffusion inversion

作　　者：李莉 张新鹏[1] 王子驰 吴德阳 吴汉舟 LI Li;ZHANG Xinpeng;WANG Zichi;WU Deyang;WU Hanzhou(Communication and Information Engineering,Shanghai University,Shanghai 200444,China)

机构地区：[1]上海大学通信与信息工程学院,上海200444

出　　处：《网络空间安全科学学报》2024年第1期92-100,共9页Journal of Cybersecurity

基　　金：国家重点研发计划(2023YFF0905000);国家自然科学基金(U23B2023,62371278,62302286,62376148);中国博士后科学基金面上项目(2023M742207)。

摘　　要：扩散模型在图像生成方面取得了显著成就,但生成的图像真假难辨,因此滥用扩散模型将引发隐私安全、法律伦理等社会问题。对生成模型的输出添加水印可以追踪生成内容版权,防止人工智能生成内容造成潜在危害。对于去噪扩散模型,在初始噪声向量中添加水印的内生水印方法可直接生成含水印图像,版权验证时通过反向扩散重建初始向量以提取水印。但扩散模型中的采样过程并不是严格可逆,重建的噪声向量与原始噪声存在较大误差,很难保证水印的准确提取。通过引入基于耦合变换的精确反向扩散,可以更加准确地重建初始噪声向量,提升水印提取的准确性。通过实验验证了引入基于耦合变换的精确反向扩散对于生成式图像内生水印的性能提升,实验结果表明,内生水印可以在生成图像中嵌入不可见水印,嵌入的水印可通过精确反向扩散被准确提取,并具有一定的稳健性。The diffusion model has achieved significant success in image generation,but it is difficult to distinguish the authentici-ty of the generated images.Therefore,abusing the diffusion model will lead to social issues such as privacy and security,legal ethics,and so on.Adding watermarks to the output of the generated model can track the copyright of the generated content and prevent poten-tial harm caused by artificial intelligence-generated content.For the diffusion model,the endogenous watermarking method of adding watermarks to the initial noise vector can directly generate watermarked images.During copyright verification,the initial vector is recon-structed through reverse diffusion to extract the watermark.However,the sampling process in the diffusion model is not strictly re-versible,and there is a significant error between the reconstructed noise vector and the original noise,making it difficult to ensure accu-rate watermark extraction.By introducing Exact Diffusion Inversion via Coupled Transformations(EDICT),the initial noise vector can be reconstructed more accurately,improving the accuracy of watermark extraction.The performance improvement of generative image endogenous watermarking by introducing EDICT has been verified through experiments.The experimental results show that endoge-nous watermarking can embed invisible watermarks in generated images,and the embedded watermarks can be accurately extracted through precise backdiffusion and have a certain degree of robustness.

关 键 词：生成式人工智能(Artificial Intelligence Generated Content AIGC)溯源 模型水印 数字水印 去噪扩散模型 反向扩散 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]