Research On Techniques Of JPEG Image Forgery Detection

Nowdays,digital images have been the main form of image storage and transmission.With the rapid development and the convenient use of visual collection devices,almost everyone has the ability to record and store digital images.In the past,people believe in the authenticity of image.However,as the rapid development of image editing software,digital image tampering becomes easier,and the tampered images are even difficult to distinguish by human eyes.Most people operate image just for beautify or entertainment,however,if someone alter images maliciously or even spread the tampered images on purpose,it will bring serious adverse effects on national security,social stability and people's normal life.Therefore,digital image forensics is of great significance.Digital image blind forensics does not rely on the image preprocessing such as signature or watermarking,which makes it wildly used and gradually become a hotspot in digital image forensics research.In this paper,the detection of synthetic tampered JPEG images are studied,which take the machine learning as the basic model.We focus on the localization of tampering regions,including double JPEG compression detection based on naive Bayesian approach,double JPEG compression detection based on a deep convolutional neural network(CNN)and inconsistency of block artifact detection based on a deep convolution neural network.The main work includes:(1)Double JPEG compression detection based on naive Bayesian approach is proposed to detect the twice compression operation in synthetic tampered JPEG images.We design two algorithms,which are based on quantization mapping relation of discrete cosine transform(DCT)coefficients,and the first statistical characteristic of DCT coefficients,respectively.For the first algorithm,by analyzing the mapping relation of DCT coefficients in the quantization process,a conditional probability model based on histogram interval mapping relation is proposed.For the second one,by analyzing the first statistical properties of DCT coefficients before and after tampering,a conditional probability model based on Benford's law is proposed.These two probability models show the changes of statistical characteristics in the process of double JPEG compression.And we locate the tampered regions combined with naive Bayesian approach.(2)Deep learning models have a strong feature learning ability and can effectively describe the intrinsic information of data.We propose a double JPEG compression detection algorithm based on a convolution neural network.The CNN is designed to classify histograms of DCT coefficients,which differ between single-compressed areas(tampered areas)and double-compressed areas(untampered areas).In this way,the most essential changes of tampering can be learned by using the CNN without complex feature design?This method can locate the tampered regions more accurately than the traditional machine learning methods.(3)The inconsistency of block artifacts detection method utilize deep learning model to detect the changes of block artifacts.Block artifacts is a weak signal which is difficult to extract,and is often impacted by the content and texture of an image.To overcome this difficulty,a simple pixel-based correlation feature combined with a deep convolutional neural network is proposed.This feature describes the effects of compression and cutting during the tampering process on block artifacts,it can study the changes of weak block artifacts in the tampering region combined with CNN.This method can locate the tampered regions,and is robust to ratating and scaling.In summary,this paper studies the method to locate the tampered regions of JPEG images on the basic model of machine learning.From the two aspects of double JPEG compression and block artifacts,methods based on naive Bayesian approach models and deep convolution neural network models are proposed.The experimental results show that the proposed methods can accurately locate the tampered region of a JPEG image.