Image Forgery Detection Using Camera Imaging Features

News photo forgery scandals emerged continuously in recent years, and thesefake photos can mislead the public and cause serious social harm. How to detectimage forgery without any prior knowledge, which is defined as passive imageforensic techniques, is a new and hot topic for multimedia security researches. Thereare a variety of existing forensic methods, which can be roughly categorized into threegroups: constraints based image content authentication methods, traces based imageforgery detection methods and intrinsic fingerprint based image source identificationmethods. In this dissertation, we mainly focus on identifying image forgery usingperspective constraints and noise feature inconsistency, and detecting the tamperingtraces left by piecewise linear contrast enhancement and image copy-movemanipulations. The contributions of this dissertation are listed as follows:1. Propose an image splicing detection method using perspective constraintsIn visual perception, distant objects appear smaller than those close to theobserver. To insert one object into an image, due to the interference of perspectiveeffects, the forger may not control the proper size of the foreign object in this imageeasily. When neglecting camera tilt and roll, via derivation, we demonstrate that theheight of any object sitting on a reference plane can be uniquely decided by thevanishing line of the reference plane, the pixel coordinates of this object in the imageand the height of camera. Furthermore, we develop this characteristic into imageforensics. To eliminate the dependence on camera height, we estimate the height ratioof two objects both resting on the reference plane and evaluate it with reference ratio.Since it do not need any prior knowledge of camera parameters, the proposed perspective constraints based method can be widely used for forensic applications.The experimental results demonstrate the efficacy of the method, even though theimages are down-sampled and high-ratio compressed.2. Develop a MAP based noise level function estimation method and apply it intoimage splicing detectionExisting noise inconsistency based image forensic methods are all with theassumption of man-made white noise. However, this assumption does not hold inmost practical sensor noise generation process. In this dissertation, the standarddeviation of noise distribution is modeled as a function with respect to image intensity,and this function is defined as noise level function (NLF). In comparison todenoising-oriented methods, incomplete sample set and unstable statistic measuringare both major challenges for application in forensics. After exposing the closerelationship between NLF and camera response function (CRF), we fit the curve ofNLF with the constraints imposed by the shape of CRF. Then we formulate aBayesian maximum a posteriori (MAP) framework to optimize the NLF estimation.Besides, we design a novel splicing forgery detection method based on the noise levelinconsistency maintaining in each image block pair from different origins.3. Propose a method to detect the trace of linear contrast enhancement andestimate the mapping parameters simultaneouslyLinear contrast enhancement is one of the common contrast modificationapproaches. Owning to the constant slope in linear mapping function, periodiccomb-like artifacts appear in image histogram. We exploit histogram frequencyanalysis method to expose this artifact. Also based on this observation, a parametersestimation approach is proposed. The peak frequency in spectrum has a strongrelationship with the slope of mapping function for estimation. To exemplify theefficacy of the proposed method, a three piecewise linear mapping function issupposed for estimation. 4. Propose an approach to detect image copy-move forgery via non-negativematrix factorization and lexicographic sorting matchingA manipulation duplicating one region of the image and pasting to another in thesame image is known as copy-move forgery. The proposed method, which belongs tothe block-feature matching category, extracts features from each block usingnon-negative matrix factorization. In order to further reduce the matching complexity,the feature coefficients are quantized to be binary and difference between each blockpair are measured by Hamming distance. A lexicographic sorting is also introduced toeliminate the redundant matchings. Low extraction complexity and high matchingefficiency are both the main advantages of the proposed method, the effectiveness ofwhich to conter filtering, compression and minor rescaling is demonstrated inexperimental results.