Scholarlink

<p>In recent years, diffusion models have exhibited remarkable capabilities in high-fidelity image generation. In this work, a robust image steganography framework is proposed, built upon diffusion models, to enable secure and controllable embedding and extraction of secret information during the generative process. To achieve this, a secret information mapping module based on orthogonal projection and local normalization is designed, allowing for efficient and stable message encoding and decoding within the latent space. In parallel, a fine-grained inversion mechanism driven by DPM-Solver++ is introduced to mitigate information degradation in the latent representation and enhance recovery robustness under adversarial perturbations. Comprehensive evaluations are conducted to assess the robustness and security of the proposed framework under various distortion conditions, including Gaussian noise, scaling, and compression. Experimental results demonstrate that the proposed method surpasses existing steganographic approaches in terms of visual quality, recovery accuracy, and resistance to detection, while maintaining high message extraction fidelity even under severe adversarial attacks.</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

Generative image steganography based on mapping-guided stable diffusion with enhanced robustness

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月5日

查看全文
  • Shengcai Zhang,
  • Junkai Fu,
  • Junxiang Xue,
  • Dezhi An

摘要

In recent years, diffusion models have exhibited remarkable capabilities in high-fidelity image generation. In this work, a robust image steganography framework is proposed, built upon diffusion models, to enable secure and controllable embedding and extraction of secret information during the generative process. To achieve this, a secret information mapping module based on orthogonal projection and local normalization is designed, allowing for efficient and stable message encoding and decoding within the latent space. In parallel, a fine-grained inversion mechanism driven by DPM-Solver++ is introduced to mitigate information degradation in the latent representation and enhance recovery robustness under adversarial perturbations. Comprehensive evaluations are conducted to assess the robustness and security of the proposed framework under various distortion conditions, including Gaussian noise, scaling, and compression. Experimental results demonstrate that the proposed method surpasses existing steganographic approaches in terms of visual quality, recovery accuracy, and resistance to detection, while maintaining high message extraction fidelity even under severe adversarial attacks.