Research On JPEG Compression Forensics

JPEG compression has been widely used in image acquisition devices,editing software and sharing platforms.The large volume and high availability of JPEG images significantly facilitates the creation of fake images for illegal purposes.Fake images can be maliciously used to mislead public opinion,interfere with judgments of courts,defraud economic benefits,or discredit government authority.In recent years,the frequently occurred crimes using fake images have presented an urgent demand for image authentication technology.JPEG compression forensics is one of the key technologies for authenticating the credibility of digital images.It can infer the compression history of an image by forensically analyzing the possible presences of JPEG compression trace,and can even locate the underlying tampered regions.After years of development,considerable encouraging progress has been achieved in the forensics of JPEG compression,but there are still many challenging issues that need to be addressed,including 1)relatively low accuracy of forensic analysis for small-sized images,and 2)lack of forensic research for JPEG-domain operations.This dissertation is devoted to addressing the above-mentioned two issues to some extent,whose contributions can be summarized as follows:1)For the forensic analysis of JPEG decompressed bitmaps,we proposed a hierarchical-clustering-based framework for improving the performance of quantization step estimation,which can remarkably improve the accuracies of the existing methods on small-sized images.The basic idea is to gather coefficient samples as many as possible by clustering different frequencies so that the insufficiency of coefficient samples at a single frequency can be alleviated.We implemented the proposed framework by using coefficient histogram and factor histogram as clustering features and conducted extensive experiments to validate its effectiveness.2)For the forensic analysis of double JPEG compression,we proposed a factorhistogram-based method for estimating the previous quantization steps of doubly compressed images.Analyzing the process of double JPEG compression,we found that the distribution of factors of coefficients can reveal the quantization step more directly than the distribution of coefficients.Inspired by this founding,we proposed the factor histogram,analyzed its statistical properties,designed the corresponding noise suppression strategy,and developed a scheme based on factor histogram for estimating the previous quantization step.Attributed to the good statistical stability and the high discriminability of factor histogram on small-sized images,the proposed method is superior to the compared methods by a clear margin in estimation accuracy.3)For the forensic analysis of JPEG-domain operations,as the first attempt in this direction,we proposed a likelihood modeling based method for the forensics of JPEG-domain enhancement.The basic idea is to transform the mixed artifacts caused by JPEG-domain enhancement to parametric JPEG artifacts by deriving the de-enhancement operator,which can be characterized more easily.Based on this idea,we constructed a novel likelihood function for de-enhanced coefficients and designed a forensic method based on the properties of the likelihood function for forensically analyzing two classical methods of JPEGdomain enhancement.Considerable experiments have demonstrated the effectiveness and superiority of the proposed forensic method.The unsatisfactory forensic performance on small-sized images and the difficulty of forensically analyzing the mixed artifacts caused by multiple operations are the challenges not only for the forensics of JPEG compression but also for other forensic tasks.This dissertation shows how these challenges can be addressed in JPEG compression forensics,which might provide some inspiration for other forensic tasks to address similar challenges.