【作者】 涂远璐；

【导师】 龚声蓉；

【作者基本信息】 苏州大学 ， 计算机应用技术， 2009， 硕士

【摘要】 隐写分析是信息隐藏检测的重要分支,分为针对性隐写分析和广义隐写分析。针对性隐写分析检测率高,但考虑到隐写术的多样性,它的实际应用受到限制。广义隐写分析适用性强,可以针对任意隐写方法训练,因此对其开展研究具有重要意义。图像广义隐写分析类似于两类模式识别问题,其目的是为了将给定图片分为隐写和未隐写两类,因此特征提取、特征空间优化和分类器设计是其中的三大关键技术,本文从这三方面进行了深入研究和大量实验,主要内容包括:1.针对目前广义隐写分析多从灰度图中提取特征而忽略了颜色信息的缺陷,在分析和比较小波子带系数高阶概率密度函数(PDF)矩和直方图特征函数(CF)矩的基础上,提出结合颜色相关度和直方图CF矩的方法。对图像亮度分量引入小波子带直方图CF矩,对色度分量提取矢量方向相关度特征。该方法不但克服了单纯从灰度图中提取特征的片面性而且通过引入直方图CF矩获得了更加有效的隐写分析特征集。实验结果表明,该算法对多种隐写方法均有较好的检测效果。2.考虑到提取的特征中有些并不能很好地反映由于隐写所带来的统计特性变化,提出利用方差分析对特征进行评估的策略,进一步从中选取更有效的特征以改善分类性能。实验结果表明,经过方差分析特征选择后,隐写分析的特征空间得到有效优化,分类器训练时间大大减少,同时检测率也得到有效提高。3.由于当前广义隐写分析多为两类隐写分析,只能判断隐写的存在性,因此通过结合直接非循环图支持向量机算法将隐写分析特征集推广到多类广义隐写分析中。这样不仅可以检测图像是否存在隐写还可以推断隐写图像可能使用的隐写工具,为进一步使用针对性隐写分析方法提供了有效信息。比较实验结果表明,该算法具有较好的分类效果,在结合广义隐写分析和针对性隐写分析方面进行了有益尝试。更多还原

【Abstract】 Steganalysis is an important branch of information hiding detection which can be divided into two categories:specific steganalysis and universal steganalysis.Specific steganalysis has higher detection accuracy but it is not practical owing to the diversity of steganography,while universal steganalysis can be trained on any Steganographic method, so it has a better adaptability.Consequently the research on universal steganalysis is a matter of great significance.Universal steganalysis is actually similar to pattern classification which centers two-class classification.The goal of it is to divide the image into two groups:cover image and stego-image.So the extraction and optimizing of features and the design of the classifier are three crucial technologies in it.This thesis carries a deep study on those three parts through a great deal of experiments and acquires a series of valuable results which can be summarized in to the following aspects:First,it is found that recent universal steganalysis schemes usually extract the features in gray images,in this way they neglect the chrominance information of images.On the basis of the comparison and analyzing between high order probability density function moments of wavelet coefficients and characteristic function moments of histogram,a universal steganalysis combining color correlation degree and characteristic function moments of histogram is proposed which extracts the features from both luminance and chrominance components of images.Moments of wavelet characteristic function is introduced to replace the wavelet high-order statistics for luminance component of image and vector directional correlation degree is calculated for chrominance component.This method overcomes the unilateralism of extracting features in grayscale images and improves the effectiveness of feature vector by introducing CF moments.The experimental results have shown that the proposed method offers a better performance over several Steganographic methods.Second,considering that some features are not able to reflect well the statistical change of the setgo-images.So a feature evaluation strategy based on ANOV(Analysis of Variance)is proposed.By exploring it,the features which are more sensitive to hidden message are selected.The results have shown that after ANOV the dimension of the feature vector reduces to thirty two and the detection accuracy has been effectively improved too.Third,take into account that most of the recent universal steganalysis schemes are two class steganalysis,so they can only judge the existence of the hidden information.Here we introduce our method into a multi-class steganalysis by using directed acyclic graph SVM(support vector machine) algorithm,This method is not only able to tell whether a image is stego or not,but also can reliably classifies it to its embedding tool.Provide effective information for specific steganalysis.The results have indicated that this method has a better classification result and makes a helpful attempt for combination of universal steganalysis and specific steganalysis.更多还原