基于多方向差分的重采样取证技术  

Resampling forensics based on multi-directional difference

作　　者：彭安杰[1,2] 曾辉[1,3] 康显桂[1,3] 

机构地区：[1]中山大学数据科学与计算机学院,广州510006 [2]西南科技大学计算机学院,绵阳621010 [3]广东省信息安全技术重点实验室,广州510006

出　　处：《中国科学：信息科学》2016年第5期627-642,共16页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:61379155;U1536204;61332012;61502547);广东省自然科学基金重点项目(批准号:s2013020012788)资助

摘　　要：重采样操作常用于数字图像篡改,重采样的盲取证受到了研究者的关注.已有的重采样取证算法主要关注取证检测器的有效性,而忽略了检测器的安全性,即恶意攻击者对检测器的攻击.目前,针对重采样取证的攻击已出现.该攻击使得基于周期性"指纹"的重采样取证检测器失效.本文提出了一种基于多方向差分的重采样取证技术.它同时考虑了重采样取证的有效性与安全性.首先根据方向性和对称性将多方向差分分组,然后分别建立自回归模型(auto-regressive model,AR)并提取出自回归系数和直方图特征,最后将所有分组特征组合成重采样检测特征.在由多个图像数据库组成的混合图像库上的测试结果表明所提出的算法既能有效地检测重采样操作,也能检测已有的恶意攻击.此外,本文提出的算法也大幅度地提升了下采样操作的检测准确率以及针对JPEG压缩的鲁棒性.篡改图像上的检测结果也证实所提算法兼顾了有效性与安全性.The forensics of resampling has recently drawn much attention since it has been used in image tampering to beautify the forged image.Existing detectors work well without malicious adversary.However,this is not possible in real life.The attack targeted against the periodicity-based method has successfully defeated the forensics of resampling.In this paper,we propose a novel detection of resampling that extracts the feature from the image difference domain.To achieve an effective and secure detector of resampling,multi-directional differences are calculated and divided into three groups.We fit each multi-directional difference into an autoregressive（AR） model.The AR parameters and histograms are averaged within each group and are concatenated to form the final feature.Experimental results on a large composite image database verify that the proposed detector is not only effective in resampling detection,but also can resist the well-known attacks.The proposed method performs much better than state-of-the-art methods,achieving significant improvement in the detection of down-sampling image and rotated image.Furthermore,the proposed method greatly enhances the detector＇s robustness against JPEG compression.The detection of tampering again verifies the effectiveness and security of the proposed feature.

关 键 词：数字图像取证 重采样 多方向差分 自回归模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]