Research On Passive Image Tampering Forensics

With the rapid development of image acquisition device and Internet technology,digital images have become our major information carrier.However,with the wide use of sophisticated and user-friendly image editing software,modifying the content of an image has become easier than before.These tampering images damage the credibility of digital media information seriously and have a negative impact in news reporting,business advocacy,academic research,and judicial judgement.Therefore,the identification of the authenticity of a digital image has become an urgent problem in the digital era.Although digital signature and watermarking,which are the representative technologies of active forensics,can be used to verify the authenticity of an image,most digital images used in our daily life do not have either.As a consequence,blind digital image forensics technology,which aims at verifying the authenticity and the originality of digital images without any a prior knowledge,has become a research hotspot.This thesis addresses the problem of blind digital image tampering forensics and identifies the authenticity of digital images from four aspects: image splicing forensics,scaling forensics,sharpening forensics,and traceability forensics.The main contributions are summarized as follows.1.For image splicing forensics,a novel image splicing detection method based on noisesharpness function is proposed.Most existing algorithms which based on noise variance inconsistencies usually ignore the influence of image content on noise variance estimation,which leads to the fact that current methods usually introduce false and missing detection on the spliced regions owning complex textures and sharp edges.For this problem,we propose a novel image splicing detection method based on noise-sharpness function.Specifically,the test image is first segmented into non-overlapping blocks and the noise variance and the sharpness of each block are calculated.Then,the noise-sharpness function,which reflects the relationship between the noise variance and the sharpness of blocks is estimated.The blocks which not constrained by the estimated noise-sharpness function can be regarded as spliced regions.Finally,contextual information is additionally used to refine the detection result.Extensive experiments demonstrate the superiority of our proposed method in detect-ing the spliced regions owning complex texture and sharp edges and is robust to additive noise to some extent.2.For image scaling forensics,a learning-to-rank approach is proposed to estimate image scaling factor.The estimation of image scaling factor can not only detect scaling operation but also help to create a complete editing history of an image.Current methods are usually unable to estimate the scaling factors of reduced images and suffer from ambiguity.For these problems,we propose a learning-to-rank approach for automatically estimating the scaling factor.With the aid of learning-to-rank,the estimation problem can be translated into classification problem.Specifically,the difference of features of the ordered image pairs are used for training and the scaling factor of a test image can be estimated from the corresponding rank values of its testing gain.Empirical experiments on extensive images with different scaling factors demonstrate the superiority of our proposed method in estimating the scaling factors of reduced images,solving the problem of ambiguity,and is robust to image size,compression quality,and additive noise to some extent.3.For image sharpening forensics,a novel image sharpening detection method based on multiresolution overshoot artifact measurement is proposed.Image sharpening,which aims to enhance the contrast of edges in an image,can be used to eliminate the tampering traces.The detection of image sharpening can serve as a reliable clue for identifying image forgery.By building the relationship between the overshoot artifact strength and the slope of an edge,we find that although undergoing the same sharpening operation,the edge with large slope will present a stronger overshoot artifact than the one with small slope.Based on this finding,we propose a novel image sharpening detection method based on multiresolution overshoot artifact measurement.Firstly,the image edge points are classified into different categories,and then the overshoot artifact strength is measured for each category respectively.Finally,a cascaded decision strategy is adopted to decide an image is sharpened or not.Experimental results demonstrate the superiority of our proposed method when compared with state-ofthe-art approaches in form of detection accuracy.Specifically,our proposed method has an obvious improvement on detecting the images underwent weak sharpening operations and is robust to image size,compression quality,and additive noise to some extent.4.For traceability forensics,an image phylogeny tree construction framework based on local inheritance relationship reconstruction is proposed.The underlying inheritance re-lationships between an image and its near duplicated images can be utilized to determine both the originality and authenticity.As the estimation of the dissimilarity matrix is usually not accurate enough,existing methods inevitably introduce local construction error during image phylogeny tree construction.For this problem,we design an image phylogeny tree construction framework based on local inheritance relationship reconstruction.We first utilize traditional approach to construct an initial image phylogeny tree and calculate the local inheritance degree of each triplet.The candidate triplets to be reconstructed can be further selected by calculating the image transformation score.Extensive experimental results demonostrate that our proposed approach can achieve better performance for image phylogeny tree construction by eliminating the local errors when compared with the state-of-the-art algorithms.