中文摘要：

针对图像隐写分析中常用的直方图特征,给出了一类基于相对熵的直方图差异计算方法,并提出了一种基于相对熵的JPEG隐写的定量分析方法.首先根据两假设检验中的最优检验——似然比检验,分析了相对熵在衡量2个直方图间的距离时的优越性,并给出了2种基于相对熵的直方图差异计算方法.然后,以新的直方图差异特征为基础,训练支持向量回归分析器,以估计隐写对DCT系数的更改比率.针对JSteg和改进的F5隐写的实验结果表明：与其他的直方图差异特征相比,根据所提出的基于相对熵的直方图差异特征所训练的定量隐写分析器具有更高的估计精度和稳定性.

英文摘要：

For the histogram features, which are usually used in image steganalysis, a category of algorithms based on relative entropy are given to measure the difference between the histograms of the given image and the estimated cover image. And a quantitative steganalysis method based on relative entropy is proposed for JPEG image steganography. Firstly, according to the likelihood ratio test, which is the optimum test for two hypotheses, the superiority of measuring the distance between the two histograms based on relative entropy is analyzed, and two algorithms are given to measure the difference between the two histograms based on relative entropy. Secondly, for the histogram-like features used in the blind steganalysis methods, the new histograms difference features are calculated by the given algorithms, and fed to the support vector regression to train the quantitative steganalyzers for JPEG image steganography. Finally, the proposed steganalysis methods are applied to the quantitative steganalysis of two categories of typical JPEG image steganography： JSteg and the improved F5 steganography. Experimental results show that compared with the other typical algorithms, the steganalyzers based on the histograms difference features calculated by the proposed algorithms have higher accuracy and stability.