Research On Blind Digital Image Forensics Methods For Authenticity Detection

With the development of modern digital technology and the availability of increasingly powerful image processing tools, digital images are easy to be manipulated without leaving obvious visual traces of having been tampered, so there is an urgent need to identify the authenticity of images. The applications of traditional digital signature and watermarking authentication technologies are limited, because they require the content providers to pre-process the images, such as extracting signature or embedding watermark. Due to the requirements of application and technology mentioned above, this thesis focuses on blind digital image forensics, which is an emerging authentication technology. As a technology of detecting image authenticity and source without relying on any pre-extracted or pre-embedded information, blind digital image forensics is becoming a new hotspot with broad prospect in the multimedia security area.This thesis makes an in-depth research on the authenticity detection problem of blind digital image forensics by applying the combined methods of theory analysis, algorithm design and experiment validation. The main contents and innovations are as follows:Firstly, the basic framework of digital image forensics is studied. As the research of blind digital image forensics technology is still in its infancy, there is no unified and mature architecture. Based on the recent developments in this field, a basic framework of digital image forensics is proposed, which consists of image modeling, feature extraction and analysis, algorithm design, test and verification, forgery area localization and image classification, image source characteristics as well as image database. The framework provides a theoretical guidance for the research of this paper, and has some reference to the future work of digital forensics.Secondly, a blind forensic approach for detecting copy-paste images is studied. The copy-paste forgery is to copy a particular part of a digital image and to cover another part of the same image. A blind image forensic algorithm based on wavelet and singular value decomposition is proposed to detect the specific forgery. In this algorithm, the copy-paste forgery detection is translated into a matching problem of similar block pairs. The wavelet transform is applied to extract the approximate component of the image, on which the sliding window operation is used. Then the singular value decomposition and quantization are adopted to extract characteristics of the fixed-size image blocks. The quantized singular value vectors are lexicographically sorted and the copy-paste forgery regions are localized by detecting all neighborhood vectors. The experimental results demonstrate that the proposed approach can detect and localize the copy-paste forgery regions accurately, and has good robustness to JPEG compression and Gaussian noise. In addition, the efficiency of our approach is improved significantly.Thirdly, a blind forensic approach for detecting inpainted image based on texture synthesis is studied. The technique of image inpainting can be used to remove objects from an image and play visual tricks. As a first attempt, a blind image forensic algorithm based on zero-connectivity feature and fuzzy membership is proposed to detect the specific forgery. Zero-connectivity labeling is applied on block pairs to yield matching degree feature of all blocks in the region of suspicion, and fuzzy memberships of these blocks are then computed by constructing a membership function. Then the tampered regions are identified by a cut set. The experimental results show that the proposed approach can effectively detect the inpainted images, which are generated by a variety of image inpainting methods, and it can localize the tampered region accurately. Furthermore, the approach has robustness to JPEG compression and Gaussian noise to a certain extent.Finally, a blind forensic approach for detecting spliced image is studied. Image splicing is a process of cropping and pasting regions from different images to form another image with necessary post-processing. A blind image forensic algorithm based on statistical characteristics of natural images is proposed to detect this specific forgery. In the algorithm, the splicing detection can be treated as a two-class pattern recognition problem. On the one hand, the generalized Gaussian distribution is adopted to model the statistical distribution of wavelet details subbands of images, and the model parameters and prediction error of each wavelet details subband are extracted as features. On the other hand, the Markov chain is applied to model the correlation of discrete cosine transform coefficients, and the state transition probability matrix is extracted as feature. Then the two kinds of features are combined to form natural image statistical feature vector, which is used to distinguish natural images from spliced images using support vector machine. The proposed algorithm is tested on three public image splicing detection databases, and the experimental results show that the algorithm has a high classification accuracy, which verifies the effectiveness of the proposed features.In conclusion, this thesis mainly focuses on the research of the systematic framework and methods for blind digital image forensics. Some achievements have been made in theory and applications. These achievements will play a positive role in promoting the development of blind multimedia forensics.