Research On Blind Forensic Methods Of Digital Image Forgery

As the image processing software and Internet improve,more and more digital tampered images emerge in people's daily life.Obtaining the processing history of digital images is an effective way to detect the authenticity of images and build a credible image system.Restoring the original information records of images,digging the potential semantics of images,protecting the security of images,and helping Judicial Forensics have great application value.Digital image forensics is a technique for analyzing and detecting digital images tampering.It aims to obtain the processing history of images.At present,digital image forensics is divided into active forensics and passive(or blind)forensics.Blind image forensic methods only need one image,without the demands of any prior knowledge and additional image samples,therefore it meets the needs of typical application,and performs automatically by software with low cost.In this dissertation,we research blind image forensic methods for image tampering.Focusing on the most common image tampering,the framework of blind image forensics for image tampering is constructed from two aspects,i.e.image compositing forensics and image retouching forensics,and a set of blind image forensic methods for image tampering are proposed.1)Research work is carried out around the image compositing forensics of the blind forensic framework for image tampering,and an image compositing forensic method based on 3D lighting model is proposed.Existing forensic methods based on3 D lighting model rely too much on reference models when reconstructing 3D models,resulting in low reconstructing accuracy and poor details,therefore the method of 3D face model reconstruction based on CNN is applied to image forensics for the first time.This method does not depend on reference models,and the reconstructed 3D model has higher accuracy and richer details,therefore it is more suitable for forensics based on lighting model.On top of that,an improved 3D lighting model is established by releasing the assumptions of constant reflectance and convex object surface,which are not suitable for face images.The experimental results show that the proposed 3D lighting model-based forensic method has higher accuracy in estimating lighting and better performance in detecting lighting inconsistency than the existing lighting model-based forensic methods.In addition,the 3D face model reconstruction method based on CNN and the released assumptions significantly improve the accuracy of lighting estimation and the performance of lighting inconsistency detection.2)Research work is carried out around the image retouching forensics of the blind forensics framework for image tampering,and a robust contrast enhancement forensic method is proposed.Foucusing on the bottleneck of existing contrast enhancement forensic methods in JPEG compression scenarios,CNN is introduced as a new method for feature extraction and learning of contrast enhancement images.A robust contrast enhancement forensic method based on CNN is proposed for the first time.In the image preprocessing progress,we design a GLCM layer to suppress the interference of image content and enhance the contrast enhancement trace.We design a cropping layer for noise reduction in the output images of GLCM layer.For the training and classification module of the network,we design tailor-made network structure and parameters for contrast enhancement forensics.The experimental results show that the performance of proposed method is still excellent in both global and local contrast enhancement scenarios,even if the image samples contained the images which were JPEG compressed and then contrast enhanced;the images which were contrast enhanced and then JPEG compressed,which make up for the blank of contrast enhancement in JPEG-compressed scenarios.The designed new layers and network structure significantly improve the performance of proposed method.In addition,the proposed contrast enhancement forensic method can be used to locate the composite image area which is retouched by contrast enhancement and therefore assists the image compositing forensics.3)Research work is carried out around the image retouching forensics of the blind forensics framework for image tampering,and a robust median filtering forensic method is proposed.Aiming at the poor robustness of existing median filtering forensic methods in detecting JPEG compressed and small-scale filtered images,image deblocking and filtering residuals fusion are used as the preprocessing methods of potentially median filtered images.Unlike existing MF forensic methods,our method begins with the analysis of the influence on median filtered images caused by JPEG compression and then effectively reduces the influence via image deblocking.On top of that,we suppress the interference caused by image content,so the fingerprints left by MF are successfully exposed.For the training and classification module of the network,we design tailor-made network structure and parameters for median filtering forensics.Experimental results demonstrate that our proposed method achieves remarkable improvements in both JPEG compressed and small size MF image detection.Furthermore,we believe that the preprocessing strategy of our method provides reference for other forensic tasks.In addition,the proposed median filtering forensic method can be used to locate the composite image area which is retouched by median filtering and therefore assists the image compositing forensics.In this dissertation,we focus on the blind forensic methods for digital image tampering,and a comprehensive framework of blind forensics for image tampering is constructed,which can be used for JPEG compressed and small size images.It provides technical support for national prosecution law and other departments,and provides scientific reference for the security and forensics of public or private digital resources.