Research On Blind Forensics Of Digital Image Tampering Based On Feature Learning

With the continuous development of information technology,digital image has become the most widely used data carrier because of its true shape,easy transmission and processing advantages.People can easily edit pictures and share them on the Internet.Once these images are maliciously falsified and spread,they may have a negative impact on the stability of society.Therefore,image forensics has become a research hotspot in the field of security,and it is of great practical significance to identify the authenticity of digital images.The research of traditional forensics methods mainly focuses on the extraction of shallow features such as frequency domain or statistics.It has some limitations,such as slow detection speed,only suitable for specific tampering operations and poor effect of image forensics after processing.Moreover,in-depth learning forgery technology is becoming more and more mature,which brings more challenges to the field of image forensics.From the perspective of feature learning,this paper discusses and analyzes the key technologies of different forgery methods,tampering with legacy features and tampering location,and carries out research by combining traditional methods and deep learning technologies.The main works are as follows:(1)A contrast tampering detection method based on second-order statistical feature learning is proposed.The existing contrast tampering detection is mainly based on the first-order histogram characteristics of the image,and cannot resist the attacks of anti-contrast forensics and other post-processing operations.This paper proposes a tampering detection method that uses an improved gray level co-occurrence matrix as the second-order statistical feature quantity and combines the active learning feature representation of the convolutional neural network.In order to learn more spatial distribution features,the feature extraction of gray level co-occurrence matrix in four directions is added,the input images with different resolutions are converted into uniform gray level co-occurrence matrix and the network structure is optimized.Experiments verify that the proposed method not only improves the detection accuracy of contrast tampering,but also can effectively detect anti-forensics attacks.(2)A splicing tampering detection method based on regional convolutional neural network is proposed.In order to realize the pixel-level localization of the tampered area,this method adds the branch of predictive segmentation mask.The edge detection operator is added to the mask branch of the regional recommendation network as an auxiliary task to increase the loss of edge supervision end,so as to encourage the predicted mask to have similar image gradient with the real mask,and learn more inconsistent features between the stitched region and the tampered region with the boundary.Experiments show that the proposed model has a significant improvement in the accuracy of splicing forgery detection and positioning.(3)An image inpainting tampering detection method based on deep convolutional neural network is proposed.This method is mainly aimed at the classic Criminisi inpainting algorithm,using the regional suggestion network as the backbone of the network,and using Soft-NMS to improve the detection rate of the candidate area frame.A large-scale image data set inpainted based on traditional and deep learning methods is constructed.In order to make full use of all levels of feature and semantic information,a multi-task repair forensic model based on the fusion feature pyramid network is further proposed.In the feature pyramid network,the reverse connection path and the texture feature extraction channel are added.The proposed method can not only detect images inpainted by traditional methods,but also has better forensic performance for inpainting based on deep learning methods.(4)A face tampering detection method based on the fusion of two-way convolutional neural network is proposed.Aiming at the forensic problem of face image tampering,combined with the advantages of dense convolutional network and deep separable network in image feature extraction,the detection rate of true and false faces is improved.In order to further obtain the location information of the face tampering,the attention module based on the spatial channel is introduced,and the attention module is proposed as the supervision information input into the two-way convolutional network to encourage the network to learn the location characteristics of the face tampering area.The experimental results prove that the proposed method can realize the localization of the face tampering area.