【作者】 汪云路；

【导师】 程义民；

【作者基本信息】 中国科学技术大学 ， 电路与系统， 2008， 博士

【摘要】 随着信息隐藏技术在网络和数字通信方面不断发展与应用,其被非法利用的危险性也相应增加。信息隐藏分析可用于网络信息的监控、分析甚至破解等,以保护网络的安全及社会的稳定。本文针对基于数字语音信息的隐藏分析作了研究,其主要内容和创新点如下:分析了信息隐藏前后语音质量统计上的变化,给出了两种语音信息隐藏盲检测方法:基于多元线性回归分类器的检测法和基于支持向量机的检测法。两者都是通过方差分析度量不同质量评估参数的可分性,选择能反映藏有信息的数字语音和未藏信息的数字语音在统计上差异的质量评估参数,用于训练学习分类器的建模和检测。前者采用多元线性回归分类器模型,后者采用支持向量机模型。实验结果表明,两种方法可以有效检测目前常见的变换域信息隐藏方法,如DFT域、DCT频、DWT域等,检测正确率均较高,使用支持向量分类器的检测效果更好。在研究质量评估参数之间相关性的基础上,提出了一种基于组合质量评估参数的语音隐藏盲检测方法。通过多维特征选取法,选择能较好区分藏有信息的数字语音和未藏信息的数字语音在统计上差异的组合参数,建立质量评估参数的支持向量机分类模型并进行检测。与不经组合特征选取的同类方法相比,提高了正确检测率,也降低了运算量。在较深入细致地研究回声隐藏方法基础上,提出了一种基于统计特征的检测策略。通过对原始语音样本和藏密语音样本复倒谱偏度和峰度的学习和训练,分别得到偏度和峰度阈值,并将其用于回声隐藏的盲检测。实验结果表明,该方法能够有效地检测出经典回声核、双极性回声核、双向回声核、双极性双向回声核等四种回声隐藏方法,运算简便,且检测率较高。设计并完成了相应的盲检测系统,进行了实验验证,取得了良好的实验结果。更多还原

【Abstract】 The development of steganography in networks and communication arises the danger for illegal activities. Stegalysis can be used for watching, analyzing and deciphering of information to protect networks and society. This paper gives some research of stegalysis based on digital audio data, and the detailed contents are as follows.After researching on the statistical discrepancy of audio quality before data embedding and after, two blind speech stegalysis methods: stegalysis based on multiple linear regression and SVM (Support Vector Machine) are proposed. They both measure quality evaluation methods depends on ANOVA (Analysis of Variance), choose some quality metrics which reflect statistical discrepancy of pure speech and stego speech well, then training classifier can be modeled. The former uses multiple linear regression classifier and latter is SVM. Simulation results indicated that they general do well in transform domain steganographic, such as DFT, DCT and DWT domain. Correct detection rate of method using SVM is higher.By analyzing the correlation of different quality metrics, an audio steganalysis based on integrated quality metrics is described. The blind detection simulating model with SVM is created with integrated audio quality metrics, which chosen by multi-dimension feature extraction, and the probably stego-audio can be detected. Simulation results indicated that multi-dimension feature extraction performs higher conect detection rate.After researching principle of echo hiding, a steganalysis for echo hiding based on statistics charcters is proposed.Coefficients of skewness and kurtosis, belonging to complex cepstrum coefficient of pure and stego speech, are compared and trained for thresholds,which used for blind steganalysis of echo hiding. Expermental results show that the presented strategy can effectively detect four kinds of echo hiding methods, including custom, positive and negative, backward and forward kernels etc.Blind stegalysis system has been designed and verified by experiment results.更多还原