Research On The Key Theories And Techniques Of Blind Image Forensics

Focusing on the topic—“research on the key theories and techniques of blind imageforensics”, we research the related theories of blind image forensics firstly, the existingtampering methods are divided into3types based on the image process, which includethe tampering of image source, the tampering of image source content and the tampering ofimage post processing. And accordingly, the basic research frame of blind image forensicsis presented. In terms of research of related technology, for the blind forensics of thetampering of image source, the blind forensics of the tampering of image content and thesystem of blind image forensics, blind identification algorithm and model are researchedfurtherly, the main research contents include the following aspects:1. Blind forensics for the tampering of image sourceIn order to improve the detection rate, we proposed a blind identification algorithm forcomputer generated images based on the discrete wavelet transform (DWT) and the mutipleangle-fractal dimension (MA-FD), which combines the statistical characteristics withgeometric characteristics of image. Firstly, the feature of transform domain of the image isextracted, which are the wavelet subband coefficients and the higher order statistics oflinear prediction errors. Secondly, we extract the geometric characteristic, which ismulti-angle fractal dimension, including global fractal dimension, the fractal dimension ofbinary images in HSV space, and the multi-fractal dimension of the density gradient imageand prediction error matrix. Finally, the SVM classifier is used to distinguish thecomputer-generated images and natural images. Experimental results show that theproposed algorithm has high detection accuracy.In addition, from the perspective of feature optimization, we present another blindidentification algorithm for computer-generated images based on LBC. Firstly, the RGBcolor space of the original image is transformed to HSV color space; Secondly, we extractthe partial binary count mode matrix of the images in HSV color space and the samplingimages, and obtain the normalized matrix histogram; Finally, the histogram above is takenas the classification feature into the SVM classifier to achieve blind identification ofcomputer-generated images. Experimental results show that the proposed algorithm reducesthe characteristic dimensions, and achieves a higher recognition rate, simultaneously.2. Blind forensics for the tampering of image contentFor the blind identification of copy-paste tampering, a blind identification algorithm was proposed based on the logarithm polar coordinate transformation for the view of imageblock matching. Firstly, the image is divided into blocks according to the gray levels ofpixels, and the suspicious blocks in the image is positioned through the gray structure;Secondly, phase correlation is made for each group of suspicious block based on logarithmpolar coordinate transformation, and the tampering regions were located roughly; Finally,tampering areas are positioned precisely. Experimental results show that the algorithm cannot only effectively detect rotated and scaled copy-paste areas, but also resist blur, addingnoise and other treatment.In addition, from the perspective of feature points matching, we proposed a blindidentification of copy-paste tampering for color images based on the SIFT and HSI model.Firstly, the SIFT key points and SIFT feature vector are extracted. Secondly, we select theHSI color model and extract HSI color information features of SIFT key points byanalyzing the advantages and disadvantages of different color space. Finally, in order toreduce the impact of non-uniform illumination, SIFT and HSI color features are normalizedrespectively, and the the most similar feature points were found by calculating theEuclidean distance. Experimental results show that the proposed algorithm can effectivelyreduce the false match rate, and has strong robustness against the Gaussian blur, white noiseand JPEG compression and other heavy post-processing operations comparing with thetraditional algorithm SIFT, SURF algorithm and improved SIFT algorithm.For the blind identification of recaptured images, we proposed a novel algorithm basedon the wavelet transform and noise analysis, through analyzing the differences of imagingprocess between the natural images and recaptured images. Firstly, we extract the meanvalue, variance and skewness of the low-frequency image and high-frequency image afterwavelet transform, which form the statistical characteristics. Then, the image noise isanalyzed by using the local binary pattern and the feature values are also extracted. Finally,the natural images and the recaptured images are classified by the support vector machine(SVM) classifier. Experimental results show that the algorithm can not only have betteridentification accuracy for the recaptured images obtained by different media, but also havelower dimension of the feature vector comparing with other exsisting algorithms.3. Blind forensics system based on the tampering evaluation modelThere are many image tampering means, but no evaluation model is designed for blindidentification system. Most researcherss select the appropriate identification algorithmbased on the artificial conjecture of image tampering means. Therefore, the intelligencelevel of blind image froensics is not high enough, and the technology does not yet form a complete and efficient blind froensic system.In order to achieve qualitative analysis of tampering means, we construct a tamperingevaluation model based on the image statistics features. The model extracts the LTP ternarypattern features, LBP texture feature and WLD local features from the image, and itconstructs multi-class support vector machine to achieve classification of the natural images,computer generated images, copy-paste images, spliced images and recaptured images. Theexperimental results show that the proposed tampering evaluation model can effectivelyclassify the tampering images in the blind environment and can objectively give the imagetampering method. Furthermore, it has laid a foundation for the subsequent authenticityidentification of the image tampering methods.Based on the tampering evaluation model above and the related blind identificationalgorithms, we construct a prototype of blind froensic system, and fulfill the systematicanalysis and identification of digital image tampering initially, which can provide areference for the study of the blind identification system.In summary, the main work of this paper focuses on the three aspects: blind forensicsfor the tampering of image source, blind forensics for the tampering of image content, andblind forensics system based on the tampering evaluation model. We have in depth studiedand discussed about analysis of the image feature, design and implementation of theidentification algorithm, the model building and other contents. We have achieved someresults in terms of the application research of theory, models, algorithms, etc., which willplay a positive role in promoting the development of information security, multimediaforensics and so on.