Research On Passive Detection Of Image Information Tampering Based On Deep Learnin

As a carrier of visual information,a digital image consists of a series of digital signals.Widely used in various electronic devices,digital images play a vital role in the progress and development of society as an important means of disseminating information.However,with advances in computer technology and digital media technology,an excessive number of image manipulation tools have emerged,which pose a threat to the security of digital image information.Image manipulation tools allow users to easily modify the original information of an image,resulting in a significant decrease in the trustworthiness of the image.The frequency of image tampering incidents has had a significant impact in the military,legal,media and other fields.As a result,this thesis analyzes and summarizes the effectiveness and shortcomings of passive image tampering detection efforts in recent years,and proposes targeted improvements in order to improve the security of image information.The main work and contributions of this thesis can be summarised as follows.(1)An image forgery detection method based on controlling learning network with multiple scales.Image information security is currently threatened by various information attacks,including image splicing forgery,which is difficult to detect due to its complex forgery method.This thesis proposes a more efficient forgery detection method based on U-Net,called Controlling Learning Network with Multiple Scales(MCNL-Net),to address the shortcomings of existing methods in terms of effectiveness and robustness.Firstly,this thesis combines the residual propagation and residual feedback modules,and inserts the merged module into U-Net to enhance its learning capability.Secondly,a multi-scale controlling strategy is proposed to design the local perceptual field of the convolutional kernel in each building block.The multi-scale structure allows MCNL-Net to learn the specified type of feature information,thereby improving its effectiveness.Furthermore,a lightweight block attention mechanism is designed to flexibly control the level of abstraction of the input feature information in each building block,thus further improving the detection performance of MCNL-Net.Experiments show that MCNL-Net can achieve more desirable results than existing methods while providing greater robustness.(2)An image forgery detection method based on multi-color space-based hybrid dense U-Net.Attackers perform tampering operations on images only in the RGB color space,therefore,existing solutions detect tampering based on this space.However,focusing on the properties of a single color space is insufficient to fully explore the tampered traces in an image.Therefore,this thesis proposes a Hybrid Dense U-Net Based on Multiple Color Spaces(HDU-Net).The thesis first analyzes and summarizes the advantages and limitations of the RGB color space,then combines RGB with YUV and LAB spaces to extract hidden feature information from the perspective of image luminance and chromaticity.Additionally,considering the sensor noise of the image itself,a steganalysis model is used to distinguish the noise difference between the tampered and authentic regions in the image,further enhancing the detection performance.The capacity of image forgery detection models is gradually increasing,and the size of the training dataset is a key factor in effectively training these models to maximize their performance.However,publicly available datasets contain limited data that cannot support the training of detection models.Thus,this thesis proposes Synthetic Adversarial Networks(SAN)to expand the amount of tampered data.By simulating the way the attacker tampers with the image,SAN can learn the correlation between scenes and objects in the image to find the most suitable hidden location to embed the tampered region in the image,enabling the synthetic image to approximate the real image.Experimental results show that SAN can expand the capacity of the original dataset by more than 40 times,allowing HDU-Net to be fully trained and significantly improving detection performance.(3)A proposed secondary labeling method for optimizing tampered data labels.Existing tampered datasets use binary annotation to label tampered and authentic regions in images.However,this annotation method only focuses on the differences between tampered and authentic regions within the same image,ignoring the differences between diverse types of tampered regions in different images.This leads to a significant loss of information contained in tampered images.Therefore,this thesis proposes a new annotation method,called Secondary Labeling(Se La).Se La restores the original differences between different tampered regions by redistributing label information to these regions,enabling the detection model to extract more hidden and effective information.Additionally,this thesis proposes an adaptive label smoothing regularization method(Adaptive Label Smoothing,ALS),which effectively solves the disassociation issue among tampered categories in Se La caused by the hardcoding method of one-hot encoding.Experimental results show that Se La can not only effectively improve the performance of the detection models,but also enhance their robustness.