Research On Blind Forensics For Digital Image Content Tampering

Nowadays, digital images are very popular in daily lives and works, and make our lives colourful. With the advent and growing popularity of sophisticated image editing software, without professional knowledge, people can perform image tampering without leaving obvious clues for human eyes. If the forgeried images are abused in news reports, courts and scientific papers, it may lead to potentially serious social, political consequences. Therefore it is very urgent to verify the authentic of digital image content.In this paper, forensic techniques of content tampering in digital images taken by digital cameras are studied. Specifically, concerning the characteristics of image storing, image tampering and its post processing, counter-methods of four kinds of common forgeries are studied:tampering of JPEG images, image splicing, region duplication and edge blurring. The main achievements are as follows:(1) A method to estimate the original quantization step of double JPEG compressed image and the corresponding approach to detect JPEG image tampering are proposed. Based on the effect of double compression on the DCT coefficients and the relationship of both quantization steps, three situations are considered. More specifically, the probability model of periodic effect is used to detect and locate the tamper region of JPEG image. Experimental results show that the proposed method can accurately estimate the original quantization step, and have good performance in locating the tampered region of the compressed image.(2) The wavelet transform with Zernike moments based tampering forensic method and Gauss Pyramid decomposition with Hu moment based copy-move method are studied. The wavelet and Gaussian Pyramid decomposition reduce the dimensions of the images, thus the dimension of feature space to be matched with are reduced. The features from the low frequency domain prove to be more robust. The method can cope with discolation of rotation attack during the tampering process by replacing square block with circle block. Simulation results show that the proposed algorithms have small computational complexity, and can successfully resist image post-processing operations.(3) A statistical model of frequency domain is proposed to detect the image splicing. The parameters of generalized Gaussian distribution model of the contourlet coefficients are estimated and the errors are calculated, the CF moments of the coefficients are also calculated. Meanwhile, the Markov transition probability matrices in DCT domain are used. Build the classifiers by combing the above features. Experimental results show that the method has good capacity in identifying the spliced image.(4) An image blur forensic method based on the edge differences is proposed. The image edges are classified according to the Non-subsampled Contourlet coefficients. Based on the differences between original and blurred image, features are extracted from image and it's predict version. Phase congruency feature is also employed, and local definition feature is introduced to tell the manual blurring from the out of focus ones. Simulation results show that the method can locate the tampered boundary with a relative high accurate rate.At last, the remaining problems of the paper are discussed, and the future works are listed.