Detecting Digital Image Tampering Using Forensic Features

Digital image is one of the most common formats to save information in moderncommunication. It is straitforward and widely used in our daily life. Since the computerscience has experienced a rapid development, one can easily change the content of the imageswhich we call ‘tampering’ via some simple softwares. It makes us hard to tell the truth and theintegrity of the images, especially when the image is cheated as evidence in the judicialforensics. So it is of great significance to analyze the algorithms of digital image forensics forprotecting the justice and maintaining social stability.From a practical point of view, we study the technical details of the forming progress ofthe digital image and investigate the characteristics introduced during this progress. At first,we introduce some knowledge of the JPEG compression and pattern noise which caused bythe sensor of the camera. Then according to the statistical properties of the images, wepropose the algorithms of digital image forensics.(1) Digital image tampering detection based on JPEG compression effectsWe firstly think that people would commonly do another different JPEG compressionafter finishing tampering a JPEG image. We analyze the effects that double compressionbrings to the histogram of the double quantized DCT coefficients of the image under threedifferent conditions. Then the tamper detection is completed by detecting the absence of thedouble compression effects through the whole image block by block. The experiments tell uswe can obtain a good detection results when the latter quantize step is smaller than the former.In view of that the man who made the tampered image would use a bigger quality factor whilethe second compression in order to cover the artifacts, so our method is fit for the actualsituation.(2) Digital image tampering detection with double check mechanism.This method is based on the pattern noise and the pattern recognition. Firstly the patternnoise of the reference camera and the test image are extracted, and then do the correlationdetection block by block. Because pattern noise we extracted would include some informationof the image content (the source of the error), we extract some features of the image blocksfor further research. We train a SVM and classify the blocks that are regarded as tamperedblocks in the correlation detection step. Finally we get the correct tampered blocks of the testimage. This method has little constraint for the format of the test image, and can obtain betterresults for different tampering means. Which is important, it has high operability and wide applicability.