Scholarlink

To address the challenges of dynamic consistency, high capacity requirements, and resistance to distortion attacks in video steganography, this paper proposes a multi-scale robust video steganography framework, RMCIN. RMCIN consists of three key modules, each designed to tackle specific challenges in video steganography. The Dual-scale Attention (DSA) module extracts cross-frame features and integrates semantic information, while the Frequency-domain Multi-scale Decomposition (FMD) module constructs a time-frequency representation through reversible transformations. Finally, the Multi-scale Fusion (MSF) module enhances the robustness of information embedding by jointly refining spatial and frequency-domain features. A specially designed Multi-step Robust Training Strategy (MRTS) progressively injects mixed distortions, including Gaussian noise and JPEG compression, dynamically strengthening the model’s resistance to attacks. Experimental results demonstrate that the proposed method achieves high-fidelity secret video reconstruction under various distortion conditions and attains state-of-the-art steganographic robustness in Gaussian noise and Joint Photographic Experts Group(JPEG) compression scenarios, offering valuable insights for real-world applications of video steganography.

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

A Robust Video Steganography Method Based on Multi-scale Decomposition and Invertible Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shiwei Li,
  • Jianxiang Liao,
  • Zhenyu Liu,
  • Mengyuan Wei,
  • Yuhang Wang,
  • Jian Liu

摘要

To address the challenges of dynamic consistency, high capacity requirements, and resistance to distortion attacks in video steganography, this paper proposes a multi-scale robust video steganography framework, RMCIN. RMCIN consists of three key modules, each designed to tackle specific challenges in video steganography. The Dual-scale Attention (DSA) module extracts cross-frame features and integrates semantic information, while the Frequency-domain Multi-scale Decomposition (FMD) module constructs a time-frequency representation through reversible transformations. Finally, the Multi-scale Fusion (MSF) module enhances the robustness of information embedding by jointly refining spatial and frequency-domain features. A specially designed Multi-step Robust Training Strategy (MRTS) progressively injects mixed distortions, including Gaussian noise and JPEG compression, dynamically strengthening the model’s resistance to attacks. Experimental results demonstrate that the proposed method achieves high-fidelity secret video reconstruction under various distortion conditions and attains state-of-the-art steganographic robustness in Gaussian noise and Joint Photographic Experts Group(JPEG) compression scenarios, offering valuable insights for real-world applications of video steganography.