计算机科学
人工智能
模式识别(心理学)
隐写术
计算机视觉
稳健性(进化)
一致性(知识库)
图像处理
图像(数学)
图像分割
离散余弦变换
噪音(视频)
特征提取
隐写分析技术
变换编码
图像压缩
算法
直方图
数据压缩
信息隐藏
数学
嵌入
作者
Guangyuan Zhang,Xuyu Xiang,Jiaohua Qin,Yun Tan
标识
DOI:10.1109/tcsvt.2026.3743946
摘要
Generative image steganography directly synthesizes stego images from secret messages without modifying existing cover images, but reliable extraction over lossy channels remains challenging for diffusion-based methods due to sampling–inversion mismatch and accumulated perturbations. To address this issue, this paper formulates robust diffusion-based generative image steganography as a closed-loop sampling–inversion consistency problem, where successful decoding requires the recovered latent variables to remain within correct decision regions. Based on this formulation, we propose a robust framework consisting of Two-Dimensional Capacity-Controllable Orthogonal Gray-code Mapping (2D-OGM), Symmetric Smoothed CFG++ Forward Sampling (SSF-CFG++), and Smoothed Fixed-Point CFG++ Inversion (SFI-CFG++), where CFG++ is a manifold-constrained variant of classifier-free guidance (CFG). 2D-OGM constructs capacity-controllable stego latents with zero-mean, unit-variance statistics and adaptive decision margins, while SSF-CFG++ and SFI-CFG++ stabilize sender-side sampling and receiver-side inversion, respectively. Experiments on multiple datasets show that the proposed method achieves high bit-level extraction accuracy under common lossy operations. Under the default 16,384-bit setting, it achieves 99.98% accuracy without attack, 98.82% under JPEG compression with quality factor 90, 99.64% under Gaussian blur with a 3 × 3 kernel, and 97.71% under resizing with factor 0.5, while maintaining competitive visual quality and limited detectability under the tested steganalyzers.
科研通智能强力驱动
Strongly Powered by AbleSci AI