Traditional methods

传统方法主要集中在对图像像素值的直接修改，以嵌入秘密信息。这些方法通常包括最低有效位（LSB）替换等技术。然而，这些方法容易被检测到，因为它们对图像的统计特征产生了显著的影响。

早期阶段

LSB

用secret messages替换最小有效位 | replaces the least significant bits with secret messages
统计特征改变 -> 导致安全问题 | indiscriminately modifying each pixel without considering the content of cover images can result in a shift in statistical features, raising security concerns.

改进的隐写编码算法

减少对像素修改的痕迹 | To mitigate the traces of pixel modifications

方案：
additive adaptive steganographic frameworks + steganographic encoding algorithms

matrix encoding
wet paper encoding
STC encoding
SPC encoding
distortion cost functions like HUGO, SUNIWARD, MiPOD, and AdaBIM

Deep Learning-based schemes

基于深度学习的方法

优点：

传统方法需要手工设计 | Traditional methods require manual design of embedding and extraction processes, making it challenging to balance distortion costs and extraction performance.
DL方法可以自动学习最优策略 | DL methods enable neural networks to autonomously learn optimal strategies.

方案：

HiDDeN pioneers the use of end-to-end training framework, realizing patch-based embedding scheme.
SteganoGAN employs generative adversarial networks to achieve high-capacity image steganography algorithms.

不足：
仍旧是modification-based方法 -> 会留下像素修改痕迹 -> 被隐写分析器(steganalyzers)感知 | However, modification-based image steganography still introduces some distortions due to the alteration of pixels. Such distortions leave traces that can be perceived by steganalyzers.

Generative Steganography

生成式隐写术

基于GAN

基于DMs

[AAAI’25] Establishing Robust Generative Image Steganography via Popular Stable Diffusion