【作者】 徐俊瑜；

【导师】 尤新刚；

【作者基本信息】 天津大学 ， 信号与信息处理， 2013， 博士

【摘要】 以数字图像、数字视频为代表的数字多媒体资源具有易编辑、易复制、易传播等特性，普通用户借助通用多媒体编辑软件便可对其进行非常逼真的编辑或篡改，且不会留下直观的视觉痕迹。数字多媒体被动取证技术就是在这种背景下应运而生的，它是在无预先嵌入特定指示性信息的情况下，“被动”地检测数字媒体的来源，及其真实性和完整性的新型认证技术。本文以数字视频资源为研究对象，围绕数字视频来源取证与典型篡改检测两个核心问题，对数字视频被动取证技术展开了较为深入的研究。主要创新性工作包括以下五个方面：(1)提出了一种基于编码开放模块差异的视频来源检测算法。通过深入分析现有视频压缩编码标准体系的特点，以视频编码标准中的两个开放模块——码率控制与运动预测为研究重点，分析并总结在MPEG-2编码标准下，不同来源的编码器之间的差异，然后基于这些差异构建了三类针对性的特征集，最后引入支持向量机实现多类视频资源的来源鉴别和追溯。(2)提出了一种针对MPEG-2标准的视频双压缩检测算法。通过定性和定量分析，发现MPEG-2标准下的数字视频在经历二次压缩编码过程后，其DCT系数分布发生规律性的变化，并以此为依据构建检测特征集，结合高性能分类算法，最终实现在多种编码参数条件下的二次压缩编码过程检测。(3)提出了一种针对MPEG-2格式转为MPEG-4格式的视频转码检测算法。在给出了重建DCT系数分布模型的基础之上，定量分析了视频转码操作对DCT系数分布造成的直接影响，并定义了全局和局部检测特征，其在不同的编码参数设置条件下，均能很好地区分原始视频资源与转码视频资源。(4)提出了一种基于高频能量变化的视频帧编辑检测算法。视频帧编辑操作会破坏原始视频帧类型分布，导致篡改视频帧序列中存在周期性的帧类型转变。同时不同编码类型帧之间对应的高频能量也因非线性量化而存在差异。本文通过构建帧高频能量特征，并引入Morlet小波分析其变化规律，在判断待测视频是否经过了帧编辑操作的同时，能够进一步探测帧编辑位置。(5)提出了一种基于频域残差的视频帧滤波检测算法。平滑滤波操作在消除噪声的同时，也会导致视频帧的高频信息大量损失，检测算法通过引入再次滤波过程，利用频域残差来分析视频帧的高频损失程度。最后，借助Radon变换与曲线建模的方法，能有效地区分原始视频帧与经过平滑处理的视频帧。更多还原

【Abstract】 Since digital multimedia resources, such as digital pictures and digital videos, areeasily manipulated, duplicated, and conveniently transmitted, the amateur caneffectively edit or tamper with digital multimedia resources by means of commonmultimedia processing and editing software without visual clues of forgery. Blinddigital media forensics technology becomes a new topic in the field of informationsecurity, which passively identifies the origination, authenticity and integrity of thedigital media without the aid of previous embedded information. This dissertationwhich belongs to the field of passive blind video forensics, focuses on two issues:source digital video identification and detection of novel tampering operation. Themain achievements reveal as following:(1) A new source video coding system identification is proposed based on thefeatures in the video stream. More specifically, it takes full advantage of thedifferent characteristics in the rate control module and the motion predictionmodule, which are two main open parts in the MPEG-2video compressionstandard. Three feature sets are extracted, and combined with a support vectormachine classifier to build an intelligent computing system for video sourceidentification.(2) A method is presented to detect double MPEG-2compression. Throughqualitatively and quantitatively analyzing the variation of DCT coefficientsduring double MPEG-2compression in depth, it is found that the distribution ofquantized DCT coefficients regularly changes. Then, a new detection algorithm isheuristically designed based on the differences in statistic distributions ofquantized DCT coefficients between the single compression and the doublecompression. Experiment results show that the proposed scheme can effectivelydetect doubly MPEG-2compressed videos under the conditions of diverse codingparameters.(3) A detection algorithm for video transcoding is proposed based on a model forreconstructed Discrete Cosine Transform (DCT) coefficients which is formulatedwith the distribution of quantization parameters. This model is utilized to revealthat the distribution of quantized DCT coefficients in the transcoded video has aseries of local maximas or minimas with a period. Finally, two sets of features are proposed to detect the periodical variation. Experimental results demonstrate thevalidity and effectiveness of the proposed approach.(4) A new method is presented to detect frame tampering based on thehigh-frequency features of reconstructed DCT coefficients in the tamperedsequences. The frame tampering operation impacts the distribution of the codingtype of frame in the original video, and then causes new periodicities in thedistribution of the coding type of frame in the tampered video. Since differentkinds of frames have different high-frequency features of DCT coefficients due tothe non-linear quantization, the distribution of high-frequency features will alsopresent periodicities. By Morlet wavelet tools, a coarse-to-fine location approachis proposed to precisely locate breakpoints in the tampered sequences.(5) A novel algorithm for detecting smoothing filtering in digital videos is proposedbased on the frequency residual. The suspected frame is re-filtered with aGaussian low-pass filter, and the difference between the initial frame and there-filtered frame in Fourier domain is called the frequency residual. Thesmoothing filtering not only eliminates the noise, but also cuts down the highfrequency information, and different filters cause different levels of distortion,which respond to different properties of the frequency residual. Finally, theRadon transform and curve modeling method are utilized to analyze thefrequency residual for distinguishing the original frame and the smoothed frame.The experimental results show that the proposed algorithm can not only detectthree typical smoothing spatial filters, including Gaussian filter, average filter,and median filter, but also can predict parameters of these filters to complementthe existing state-of-the-art methods.更多还原