Abstract: In order to solve the problem that the majority of general steganalysis methods cannot estimate the secret message length, this paper proposes an improved general quantitative steg-analysis method that can estimate secret message length. 132 dimensional features describing the correlations between DCT coefficients are extracted from stego images. Support vector regression is used to learn the mapping between feature vectors and the relative embedding change rates and construct steganalyzer model. Embedding rates are estimated through new feature sets and steganalyzer model. Simulation is performed on stego images embedded with F5 , MB and outguess steganographic algorithms. The results of simulation reveal that the proposed quantitative steganalysis is feasible to estimate the embedding ratio of stego images in practice.

Key words: quantitative steganalysis, support vector regression, loss function, kernel function

摘要： 为解决大多数通用隐写分析算法不能检测秘密信息长度的问题，提出了一种改进的能估计秘密信息长度的通用隐写分析方法。从隐写图中提取描述DCT域系数相关性的132维特征，用支持向量回归机学习图像特征和相应嵌入改变率之间的映射关系并建立模型，根据映射模型估计测试隐写图的嵌入改变率。使用典型的嵌入算法：F5、outguess与MB进行测验，仿真结果显示提出的秘密信息长度估计算法是切实可行的。

关键词: 通用隐写分析, 支持向量回归, 损失函数, 核函数