数字图象被动取证中若干关键技术的研究

Passive digital image forensics, a technology for image authentication and source identification without the aid of any pre-extraction or pre-embedding information, is beneficial in many practical applications where the original image is not directly accessible, and is emerging as a hot topic in the community of multimedia security. Both resampling detection and median filtering detection, which relate to each other, have become two standard tools for the forensic analysis of digital images. The creation of convincing image forgery, e.g. copy-move or image splicing, oftentimes involves geometric transformations, which typically comprise a resampling of the original image to a new sampling grid. Therefore, the detection of resampling traces reveals the possible tampering regions in the image. The median filtering, however, serves as an effective counterforensics tool for hiding traces of resampling, due to its non-linearity and complex statistical characteristics. Consequently, the median filtering detector is also an essential forensic tool for exposing possible counterforensics operations. This project investigates some key technical issues involved in the theory and practice of both resampling detection and median filtering detection, which includes: 1) Design and implementation of robust affine transformation detection and estimation, by incorporating the two-dimensional cyclostationary characterization of resampled image with affine transformation. 2) Analyze and analytically characterize the statistical artifacts brought into the difference domain of images by median filtering (MF), through deriving several key formulas of cumulative distribution function (CDF) in the difference domain of different image sources, such as median filtered image, average filtered image and original non-filtered image. 3) Design and implementation of new median filtering detection scheme based on the statistics in difference domain, which is effective and reliable for both uncompressed and JPEG post-compressed images, even in the cases of low resolution and deep JPEG compression.

作为一种不依赖任何预签名提取实现数字图像真伪和来源鉴别的技术，数字图像被动取证具有更广泛的适用性，正日益成为多媒体安全领域的研究热点。重采样检测和中值滤波检测是当前数字图像被动取证中两种重要且相互关联的分析工具；一方面，通过对伴随图像局部复制-粘贴或图像拼接操作产生的重采样痕迹的检测可以揭示图像篡改区域的存在；另一方面，由于中值滤波可以有效去除图像中的重采样痕迹，对中值滤波的检测又可以揭示图像中可能的反取证操作。本项目研究重采样检测和中值滤波检测理论和方法中的几个关键问题，包括：（1）基于仿射变换重采样图像的二维循环平稳表征，研究鲁棒的仿射变换图像检测和仿射矩阵估计方法；（2）通过推导中值滤波图像和其他平滑滤波图像差分域累积分布函数（CDF）的分析表达式，研究对相关图像操作的可区分统计表征；（3）研究和发展基于图像差分域统计特征的中值滤波检测新方法，实现对数字图像中值滤波的鲁棒检测。

作为一种不依赖任何预签名提取或预嵌入信息实现数字图像真伪和来源鉴别的技术，数字图像被动取证具有更广泛的适用性，正日益成为多媒体安全领域的研究热点。重采样检测和中值滤波检测是当前数字图像被动取证中两种重要且相互关联的分析工具；一方面，通过对伴随图像局部复制-粘贴或图像拼接操作产生的重采样痕迹的检测可以揭示图像篡改区域的存在；另一方面，由于中值滤波可以有效去除图像中的重采样痕迹，对中值滤波的检测又可以揭示图像中可能的反取证操作。本项目主要研究重采样检测和中值滤波检测理论和方法中的几个关键问题，包括：（1）基于仿射变换重采样图像的二维循环平稳表征，研究鲁棒的仿射变换图像检测和仿射矩阵估计方法；（2）通过推导中值滤波和其他平滑滤波图像差分域累积分布函数的分析表达式，研究和实现对相关图像操作的统计表征；（3）研究和发展基于图像差分域统计特征的中值滤波检测新方法。在项目研究过程中，研究团队针对以上内容开展了卓有成效的研究工作，对于连续图像重采样检测，首次从理论上给出了连续仿射变换图像频域特征极点位置和连续仿射变换矩阵之间的关系式，并据此提出和实现了一种针对连续旋转或尺度拉伸图像的有效盲取证方法，解决了多媒体内容安全领域自2010以来的一个“公开”问题。与此同时，研究团队通过研究中值滤波图像差分域统计分布的表征，提出和实现了一种鲁棒的高性能中值滤波图像的盲检测方法，并被同行公认为当前几种性能最好的图像中值滤波检测方法之一。在项目执行期，研究团队也在多媒体内容安全的其他研究方向取得了显著进展，包括图像隐写和隐写分析、图像可逆信息隐藏、图像拼接检测和加密域图像压缩等，共计发表学术论文近30篇，其中包括10篇IEEE会刊论文，较好地完成了本项目规定的研究任务。