Research On Passive Forensics Of Digital Image Authenticity

Passive forensics of digital image authenticity refers to the technology to make judgment whether an image is a photographic image generated from a digital camera, or whether the content of the image has been manipulated, without any priori knowledge but with certain technical approaches. The main purposes of digital image authenticity forensics include the analysis and forensics about image source, undergone manipulations of the image, and the authenticity of the image content. The technology of digital image authenticity forensics can provide technical support to meet the requirement of digital image authenticity detection in various areas such as journalism, publishing, justice and social media, and effectively counter and overpower various malicious image manipulations.From the standpoint of digital image authenticity forensics, the research content in this thesis mainly includes distinguishing photorealistic computer graphics from natural images, JPEG compression history detection and doctored image detection. The main research results of this thesis include:1. Distinguishing photographic images and photorealistic computer graphics(1) Based on the existing approaches to image identifying depending on the statistical characteristics of wavelet coefficiences, imaging features and visual features of images are studied. Motivated by wavelet based image denoising, an image is divided into subimages corresponding to image detail and noise information. Then, statistical features are extracted from different components. The SVM classifier is employed to identify photographic images and photorealistic computer graphics.(2) Based on the analysis of statistical properties of geometrical structure of local image edge patches, two features are presented for identifying photographic images and photorealistic computer graphics. First, 3×3 image edge patches are projected into a 7-dimensional sphere and the 7-dimensional sphere is divided into 17,520 Voronoi cells. Consequently, the distribution of data points on the 7-sphere is converted to calculate the probabilities that data points fall into the corresponding Voronoi tessellations. One of the proposed features characterizes the structure difference of local patches between the two image categories. A test image is identified by computing the similarity between it and the image models. Considering local edge patches of the two image categories with different response to compression manipulation, one more feature is presented to capture their different change before and after compression. Experimental results demonstrate that the two features have efficient discrimination ability on various datasets.(3) Based on the sparseness property of the geometrical structure of 3×3 local image edge patches, an approach based on visual vocabulary to distinguish photographic images and photorealistic computer graphics is proposed. First, we propose a novel approach to constructing visual vocabulary with the key sampling points corresponding to specific Voronoi cells, which avoids some fatal problems of the previous bag-of-words model. Based on the constructed visual edge vocabulary, images are classified by employing the SVM classifier. Experimental results demonstrated that the visual-vocabulary based method has better efficient discrimination of the two image categories and anti-compression ability.(4) An approach based on the histogram of tetrolet covering(HoC) to distinguishing photographic images and photorealistic computer graphics is proposed. The optimal tetrolet covering of a 4×4 image block decribes the local geometry structure. Utilizing the adaptive approach of Tetrolet transform to search the optimal tetrolet covering, we analyze the statistical characteristics through the the histogram of tetrolet covering. The experimental results demonstrate the HoC feature extracted from S(saturation) component outperforms the existing approaches to identify photographic images and photorealistic computer graphics and to resist JPEG compression. This ensures it is promising to distinguish photographic images and photorealistic computer graphics oriented to mass image library.2. JPEG compression history detection(1) JPEG compression history detection based on dqcurves. We run a recompression operator on the test image. We measure the JPEG quantization error by computing the sum total of square difference between the given image and the recompressed version, and the error is normalized and denoted as dq. We search the local minima on the dqcurves in the Y, Cb and Cr color channels and determine the compression history of the image, including compression times, compression sequence with the quality factors. Moreover, we put forward a specific application of the compression history detection, i.e., in-camera compression detection of a camera image, by which the source of a camera image can be authenticated.(2) JPEG compression history detection based on tetrolet covering change rate. Utilizing tetrolet covering, we convert 4×4 image block in a color channel from the spatial domain to the tetrolet-covering domain. We compute the tetrolet-covering change rate of the image under compressions of different quality factors and obtain corresponding tetrolet-covering change rate curves. Likewise, the compression history of the image can be exposed by detecting local minima on the curves. Algorithms of different compression tools and multiple-compression detection are described. Both the proposed detection approaches run on a single image, do not need complicated classifier design and a large number of training samples, and then avoid the model training process of the machine learning method. They are simple, reliable and efficient. So far, most existing approaches to detecting JPEG compression history are only able to detect single or double compression, and very little has been reported about the approaches of in-camera compression detection and different compression tools and multiple-compression detection.3. Exposing composite image by detecting inconsistences in compression historyBesides the research on tetrolet-covering based compression history detection, tetroletcovering map is further introduced to expose composite image. The inconsistences in compression history are visually presented by the change of tetrolet covering index under a recompression. According to the drawn conclusion that with the increase of compression times, quantization effect becomes smaller, the tampered image parts with more compressions and the original image parts present inconsistence in the change rate of tetrolet covering index. This helps us find the tampered part of a composite image. As we observe the change of an image block due to a recompression depending on tetrolet covering, this helps this approach to overcome the misalignment of JPEG lattices when the images are splice. Compared with the existing pixel based approach, the proposed tetrolet-covering based approach has its strength.