Research On Forgery Detection In Image Operator Chains

With the emergence of numerous emerging social media platforms,digital images have become an essential carrier for social media to convey information due to their intuitive and easy-to-understand characteristics.However,the popularity of image editing tools has made image forgery increasingly easy.The spread of fake images often leads to public misconceptions,which harms politics,economics,social culture,and other areas.Therefore,studying image authenticity detection is of great significance.Currently,most image forensics techniques are aimed at detecting specific tampering operations.However,image forgery usually involves multiple operations,which constitute an image operator chain in a specific order.Although some initial progress has been made in image operator chain forensics,the accuracy of operator chain detection,the generalization of forgery localization across database,and the robustness against lossy operations still need to be improved.To address these challenges,this thesis will focus on the type and order detection of operations in the global tampering operator chain and forgery localization in the semantic tampering operator chain,specifically including the following four works:(1)Research on global tampering operator chain detection based on operational feature decouplingAn operational feature decoupling method is proposed to solve the problem that tampering traces are difficult to detect due to the superimposed processing artifacts caused by multiple tampering operations.By analyzing the relationship between blind signal separation and feature decoupling,it is found that it is reasonable to use the feature decoupling method based on blind signal separation to realize operator chain detection.This method estimates the source features of each operation from coupled features and uses them as effective evidence for global tampering operator chain detection.In addition,a tampering operator chain detection strategy based on decoupled features is proposed to realize operation type and order detection.Finally,through comparison with existing image operator chain forensics methods,it is shown that the detection accuracy in different image operator chains by using the proposed method is 18% higher than that of the existing methods.(2)Research on global tampering operator chain detection based on trustworthy decision fusionA trustworthy decision fusion method is proposed to solve the problem of performance degradation of single tampering operation forensics methods in detecting tampering operator chains.This method allows the integration of forensic knowledge provided by multiple existing image forensics algorithms.Specifically,this method first introduces a similarity coefficient concept to allocate credible weights for forensic evidence by calculating the similarity between different forensic knowledge.Secondly,multiple pieces of forensic evidence are integrated using fusion rule based on local conflict handling.Then,based on trustworthy decision fusion,a tampering operator chain detection strategy is proposed to identify image tampering history under different scenarios.Finally,effectiveness verification was carried out in different global tampering operator chains.The results show that the accuracy of the proposed method in identifying image tampering history is 26% higher than that of the existing operator chain forensics methods.(3)Research on post-processed image forgery localization based on local signal noise separationA local signal noise separation-based network for image forgery localization is proposed to solve the problem that existing image forgery localization methods are not robust to post-processing attacks and have weak cross-database generalization.The network first uses a signal noise separation module to separate semantic tampered areas from complex background areas with post-processing noise,weakening or even eliminating the negative impact of post-processing operations on forgery localization.Then,high-level global features are learned from multiple perspectives using a multiscale feature learning module based on parallel dilated convolution architecture.In addition,by using a feature fusion module,the discriminative ability of the tampered area is enhanced by strengthening boundary information.Extensive experiments show that the proposed network is robust against various post-processing attacks and the localization accuracy of semantic tampered regions is improved by 11% compared with the state-of-the-art forgery localization methods.Meanwhile,the proposed network has generalization for detecting forged images from unknown sources.(4)Research on social media shared image forgery localization based on tampering probability estimationA tampering probability estimation-based network for image forgery localization is proposed to address the problem of the reduced forgery detection performance of existing image forensics methods due to lossy operations performed by online social media on uploaded images.The network first uses a probability estimation method based on feature coupling and selective search to obtain the tampering probability map of the local regions.By mining forgery traces in the spatial domain,residual domain and wavelet domain,lossy noise introduced by transmission on social media can be suppressed,and the most differentiated regional information in images can be captured.Furthermore,a forgery detector based on low-level and high-level feature learning is proposed to adaptively capture more detailed local edge differences between tampered and real areas.Extensive experiments show that the proposed network can robustly expose the semantic tampered regions of forged images transmitted on different online social media in real time.Meanwhile,in the forgery detection of image databases containing multiple semantic tampering operations,the accuracy of the proposed network is 6.8% higher than that of the state-of-the-art forgery localization methods.To sum up,this thesis proposes corresponding solutions to the main challenges faced by image operator chain detection.Extensive experiments have verified the effectiveness of the proposed methods in improving the accuracy,generalization and robustness of operator chain detection.Thus,this thesis will provide more effective technical support for the authenticity of judicial electronic image evidence and image news report.