The Research On Double JPEG Compression Forensics Based On Convolutional Neural Networks

Since it can maintain good visual quality of an image while gets rid of much redundant information,JPEG image coding technique is widely used to save the transmission bandwidth and storage space.Most image processes such as Internet transmission or camera internal storage adopts the technique.Because of this widely usage,JPEG images are vulnerable to the attacks,which can be some editions to improve aesthetic effects such as contrast enhancement,and also can be the vicious modifications to deceive the masses or hide the evidence,such as to wipe some contents of an image or copy-move the target to an image.Therefore,in some specific occasions,to verify the authenticity and originality of an image is very important and necessary.Image forensics is a series of the digital techniques applied to these scenarios.Double JPEG compression detection has great significance to image forensics,for the possibility that it can reveal whether an image is tampered and possibly further to locate the tampered regions.Many researchers study the JPEG image characteristic with respect to DCT-domain and spatial-domain knowledge,to find out that whether the image is JPEG compressed twice.Recently,convolutional neural network(CNN)has gained a lot of success on the task of image classification and text recognition.This leads to that a few researchers have introduced CNN on detecting double JPEG compression.However,those existing approaches have some defects.For the handcrafted features based approaches,they focus on some kind of statistical property,for example,the distribution of DCT coefficients,which could be cover by the anti-forensics experts.Then for the existing CNN based approaches,they more or less already make some extracted features as input,and moreover,the networks do not completely utilize the JPEG image characteristic information.The major task of this dissertation is to comprehensively consider and analyze the characteristics of a JPEG image,exploring the model based on convolutional neural network to improve the detecting accuracy of double JPEG compression.We present: 1)a multi-branch DCT-domain based convolutional neural network,JPEG-CNN,to learn the information intra-sub-band and across sub-bands;2)a double-access spatial-domain based convolutional neural network,SB-CNN,aiming at the spatial blocking artifacts of JPEG decompressed images;3)a modified DCT-domain based neural network,DC-CNN,and a double-domain based convolutional neural network is built.It adopts information from different domains simultaneously to enhance the learning ability.The experimental results show that:(1)JPEG-CNN has relatively strong capability to detect double JPEG compression,and it is comparative with traditional hand-crafted features based approaches.Moreover,it has made an improvement with 4.61% in individual QF combination.(2)SB-CNN has slightly lower ability than DC-CNN on detecting double JPEG compression,but also maintains in the same level compared to hand-crafted features based approaches and CNN based methods.However,SB-CNN shows more powerful capacity when detecting the double JPEG compressed images received anti-forensics attacks,making an improvement from 9% to 40%.(3)Modified JPEG-CNN,namely,DC-CNN,has steady and improved performance of detecting double JPEG compression and those received anti-forensics attacks,respectively.And from overall outlook,the combined network DD-CNN has both surpassing effectiveness and robustness.At the same time,we demonstrate comprehensive experimental results,implemented with multiple anti-forensics methods,forensics techniques and way of classification.It indicates that,our DD-CNN is more stable and reliable than other compared approaches.