| 引用本文： |
-
白华元,符皓程,张弘,曹纭.最小化像素分布失真的重采样式自适应图像隐写方法[J].信息安全学报,已采用 [点击复制]
- Bai Huayuan,Fu Haocheng,Zhang Hong,Cao Yun.A Resampling-Style Adaptive Image Steganography Method with Minimized Pixel Distribution Distortion[J].Journal of Cyber Security,Accept [点击复制]
|
|
| 摘要: |
| 近年来，基于自适应隐写嵌入的修改式图像隐写方法正逐步从依赖启发式代价函数的传统方法，向具有更强数学理论基础的、基于图像分布建模的隐写范式演进。与此同时，生成式隐写的发展为本研究提供了重要启示。本文融合自适应隐写框架的可靠性与生成模型在图像分布建模方面的优势，提出一种旨在最小化像素分布失真的重采样式自适应图像隐写方法。首先，构建基于图像恢复模型的像素分布估计器：通过对参考图像施加微小扰动，并借助恢复模型生成多个内容语义一致但细节不同的图像样本，进而估计各像素位置的条件高斯分布，实现对图像局部像素分布的高精度建模；其次，设计一种基于拒绝采样的重采样式嵌入策略，将秘密信息嵌入量化后的像素取值中，在不显式更改原始像素值的前提下完成信息隐藏，从而实现“分布内采样”式的隐写机制；最后，提出一种闭式表达的分布失真代价函数，以隐写前后像素分布之间的KL散度作为优化目标，并结合STCs等高性能的隐写编码技术，实现自适应的信息嵌入。实验结果表明，所提出方法在不同嵌入率下生成的隐写图像在视觉质量上与自然采样的图像难以区分，且在统计不可区分性方面显著优于当前主流的同类图像隐写方法。 |
| 关键词: 图像隐写 图像恢复模型 拒绝采样 KL 散度 多元高斯载体模型 |
| DOI： |
| 投稿时间：2025-08-01修订日期：2025-11-18 |
| 基金项目:国家自然科学基金,国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| A Resampling-Style Adaptive Image Steganography Method with Minimized Pixel Distribution Distortion |
|
Bai Huayuan, Fu Haocheng, Zhang Hong, Cao Yun
|
| (Institute of Information Engineering, Chinese Academy of Sciences) |
| Abstract: |
| In recent years, modified image steganography based on adaptive steganographic embedding has been progressively evolving from traditional methods relying on heuristic cost functions toward a more mathematically grounded paradigm that leverages image distribution modeling. Meanwhile, advances in generative steganography have provided significant in-spiration for this work. By integrating the reliability of adaptive steganographic frameworks with the strengths of generative models in distribution modeling, this paper proposes a resampling-style adaptive image steganographic method aimed at minimizing pixel-level distribution distortion. First, a pixel distribution estimator based on an image restoration model is constructed: by applying small perturbations to a reference image and leveraging the restoration model to generate multiple image samples with consistent semantic content but diverse fine-grained details, the conditional Gaussian distribution at each pixel location is estimated, enabling high-precision modeling of local pixel distributions. Second, a resampling-style embedding strategy based on rejection sampling is designed, wherein secret information is embedded into quantized pixel values without explicitly altering the original pixel intensities, thereby achieving a "within-distribution sampling" ste-ganographic mechanism. Finally, a closed-form distribution distortion cost function is introduced, with the KL divergence between pixel distributions before and after embedding serving as the optimization objective. This cost function is com-bined with high-performance steganographic coding techniques such as STCs to realize adaptive information embedding. Experimental results demonstrate that the stego images generated by the proposed method are visually indistinguishable from naturally sampled images under various embedding rates, and exhibit significantly enhanced statistical indistin-guishability compared to state-of-the-art image steganographic methods. |
| Key words: image steganography image restoration model rejection sampling KL divergence multivariate Gaussian cover |
