Research On Image Order Forensics For Multiple Manipulation Operations Based On CNN

With the rapid spread of various digital multimedia devices,digital media has gradually become a main information carrier.However,the manipulated multimedia data has also led to various risks of information security.Image passive forensics is a research of analyzing the inherent features of content of digital images with the aim of detecting the authenticity and integrity of testing images.And it's one of the research focus in the field of information security.However,the research of the traditional digital image passive forensics technology mostly focus on single tampering operation,and the performance strongly depends on the quality of the artificial features that needs the professional knowledge of the forensic domain.The convolutional neural network can automatically learn features from the image and identify them,and it reduces the dependence on the knowledge of the forensic domain.In this paper,we study the resear ch of forensics of image multiple consecutive manipulation based on convolutional neural networks,and the main results are as follows:(1)A general image forensics framework based on convolutional neural network are designed for multiple consecutive operation order forensics.The forensic framework contains two sub-streams,the spatial convolution stream can extract the visual tampering feature from image spatial domain,and the transfer-feature extraction stream can extract the transform residual features of image transform domain with using a well-designed processing.And the forensic framework also fuses the transform residual features and image spatial features together to generate the satisfying classfication results.In addition,we also adopt transfer learning strategy to improve the detection performance of some low-intensity tampering operation chains that is hard to detect.By considering the common pattern among different tampering intensity under the same one tampering operaiton,we adopt the transfer learning strategy to fine-tune the training to improve the order detection performance of the low intensity tampering operation chain.The experimental results show that both the proposed order forensics framework based on convolutional neural network and the transfer learning strategy can improve the performance of order forensics for operation chain.(2)In order to detect the order of tampering operation chain for manipulated JPEG image,we utilize DCTR-based processing to realize the DCT orthonormal projection of spatial feature in JPEG image.And for tampering operation with unknown intensity,we construct a tampered image dataset generated by various tampering parameters to extract common traces of the same tampering operation under different intensities,and realize order forensics of the operation chain without knowing the specific tampering parameters.The experimental results show that DCTR-based preprocessing can effectively improve the order detection performance of JPEG images,and the network model trained on the image dataset that contains various tampering parameter is robust.