Research On Image Tampering Forensics Based On Deep Learning

With the development of image processing technology and the popularity of image editing software,it has become easier to modify an image.The malicious spread of tampering images on the Internet will have an impact on social stability,public credibility and media value,and thus digital image forensics technology for identifying image authenticity and integrity has received increasing attention.The key to digital image forensics research is to be able to design tampering features that meet the purpose of forensics.However,manual design features are difficult and hard to apply universally.In recent years,deep learning,especially the widespread application of convolutional neural networks in the field of computer vision has brought new solutions to the problem of image tampering forensics.Convolutional neural networks can automatically learn data distribution from datasets by supervising learning,and apply data distribution to specific tasks.From the perspective of feature learning,this paper uses the powerful feature extraction ability of convolutional neural networks to improve some existing classical network models and apply them in the field of image forensics.Aiming at the problem of image manipulation forensics,this paper proposes an image manipulation forensics method based on deep residual network.The method firstly uses the high-pass filter pre-processing layer to extract the tampering noise residual signal,and then uses the residual module to perform layer-by-layer feature extraction and fusion,and finally integrates the features into the sample marker space.The proposed network model can be used for detection of five image manipulation types such as Gaussian blur,sharpening,median filtering,JPEG compression,and Gaussian white noise.The mean accuracy rate is 98.27% on the ImageNet224 dataset.At the same time,the network also has good forensic performance for small size images.Aiming at the problem of image tampering area localization,this paper proposes a tampering area localization method based on fully convolutional network and conditional random field.The method firstly trains three different scale fully convolutional networks for rough extraction of tampering regions,then fuses the localization results of the three networks,and then processes the localization results with a conditional random fields to correct the relationship of neighboring pixels.Finally,the final localization result is determined by the threshold.Compared with several FCN-based tampering area localization methods,the designed network has good detection performance on the CASIA ITDE dataset.At the same time,the algorithm also has good detection results for the splicing images with image manipulations which shows great robustness.