Scholarlink

<p>To address issues of limited watermark information for copyright traceability and weak resistance to reconstruction de-watermarking attacks, this paper proposes a robust and large-scale traceable watermarking method for diffusion models using accurate encoding of affine coupled flow. The watermark encoder and decoder are constructed using affine coupled flows. The encoder embeds watermarks into latent variables following a standard normal distribution, consistent with diffusion model priors. Reversible flow models enable precise watermark encoding and extraction in latent space, allowing semantic-level embedding without compromising image quality. Fine-tuning with an attack layer and adversarial training enhances decoder robustness. The method generates high-quality watermarked images in real time by encoding user-specific information as hidden variables input into a pre-trained diffusion model, requiring no additional training. Experiments demonstrate over 97% detection and bit accuracy under image processing and reconstruction attacks, outperforming leading methods. In large-scale user scenarios, watermark traceability remains at 95% under image processing attacks, ensuring reliable copyright protection.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Large-scale traceable and robust watermarking for diffusion models with precise affine coupling flow encoding

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Jiayue Zhang,
  • Xiaobing Kang,
  • Yiting Guan,
  • Jiayi Du

摘要

To address issues of limited watermark information for copyright traceability and weak resistance to reconstruction de-watermarking attacks, this paper proposes a robust and large-scale traceable watermarking method for diffusion models using accurate encoding of affine coupled flow. The watermark encoder and decoder are constructed using affine coupled flows. The encoder embeds watermarks into latent variables following a standard normal distribution, consistent with diffusion model priors. Reversible flow models enable precise watermark encoding and extraction in latent space, allowing semantic-level embedding without compromising image quality. Fine-tuning with an attack layer and adversarial training enhances decoder robustness. The method generates high-quality watermarked images in real time by encoding user-specific information as hidden variables input into a pre-trained diffusion model, requiring no additional training. Experiments demonstrate over 97% detection and bit accuracy under image processing and reconstruction attacks, outperforming leading methods. In large-scale user scenarios, watermark traceability remains at 95% under image processing attacks, ensuring reliable copyright protection.