Scholarlink

<p>Generative steganography has gained significant attention due to its cover-independent nature, which allows it to evade a broad range of steganalysis techniques. However, existing methods often suffer from poor extraction accuracy at high embedding rates and limited robustness to real-world noise. To address these limitations, we propose the Controllable Generative Image Steganography (CGIS) framework, introducing two bidirectional mapping mechanisms: the Resampling-based RS mapping and the Normal Distribution-based ND mapping. These significantly enhance extraction accuracy even at high embedding rates. Unlike prior stochastic or training-dependent approaches, CGIS utilizes a training-free and deterministic diffusion process, fundamentally distinguishing it from existing methods and ensuring superior robustness. Additionally, a Gaussian Mixture Model-based image denoiser (GMM-ID) ensures the model’s resilience to network noise and compression-induced distortion. Experimental results demonstrate that CGIS outperforms existing methods in terms of extraction accuracy (97.75%), detection resistance (96.23%), and image distortion (FID: 2.36), demonstrating superior performance for practical deployment.</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

Controllable generative image steganography based on denoising diffusion implicit model

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月10日

查看全文
  • Tian Wu,
  • Xuan Hu,
  • Chunnian Liu,
  • Ting Liu,
  • Lina Wang

摘要

Generative steganography has gained significant attention due to its cover-independent nature, which allows it to evade a broad range of steganalysis techniques. However, existing methods often suffer from poor extraction accuracy at high embedding rates and limited robustness to real-world noise. To address these limitations, we propose the Controllable Generative Image Steganography (CGIS) framework, introducing two bidirectional mapping mechanisms: the Resampling-based RS mapping and the Normal Distribution-based ND mapping. These significantly enhance extraction accuracy even at high embedding rates. Unlike prior stochastic or training-dependent approaches, CGIS utilizes a training-free and deterministic diffusion process, fundamentally distinguishing it from existing methods and ensuring superior robustness. Additionally, a Gaussian Mixture Model-based image denoiser (GMM-ID) ensures the model’s resilience to network noise and compression-induced distortion. Experimental results demonstrate that CGIS outperforms existing methods in terms of extraction accuracy (97.75%), detection resistance (96.23%), and image distortion (FID: 2.36), demonstrating superior performance for practical deployment.