Research On Key Technologies Of Digital Imaging Evidence Oriented Forensic Inspection

In recent years,image has become an important carrier for disseminating and preserving imformation as acommon data format.In both electronic devices and network environment,images can reflect more user information than other data types.The authenticity and primitiveness play important roles in many fields,such as criminal investigation,image forensics,news media and military investigation.The traditional analysis and processing methods expose some problems in the acquisition and the identification of imaging evidence as the volume of imaging data increasing rapidly,and the forgery techniques being developed daily.Under the background of the artificial intelligence,the key to promote the development of digital image forensics is to investigating forensic inspection technology of digital images,combining advanced technologies of computer vision to improve the efficiency and performance.This research will promote the development of image forensics,and is very helpful for protecting the security of information and cyber space.This paper focuses on the research of the key technologies of forensic analysis of imaging evidence in digital forensics.According to the relevant regulations and requirements of digital forensics and the related technologies in the field of computer vision,a typical model of forensic analysis oriented to imaging evidence is designed and put forward,which mainly guidence three functions: forensic recognition of digital images,tampering detection and source identification of digital images.Then,specific methods and technologies are developed for the above three functions to improve the efficiency of digital image forensics,to obtain useful information in massive raw data,and to test and identify whether the evidence meets the requirements of relevant laws and regulations for authenticity,relevance and primitiveness.The main research work and innovative contributions of this paper is summarized as follows:(1)A typical model for forensic inspecting of digital imaging evidence is proposed.This model ensembles advanced technology of image recognition,image tampering detection and others.Its framework covers the whole life circle of digital imaging evidence,including module of feature extraction,module of forensic recognition,module of tampering detection,and module of source identification.The feature fusion module serves as the basis for subsequent forensic analysis.The forensic recognition module is responsible for the identification of key evidence,while the tampering detection and source identification modules verifie the authenticity,relevance and legitimacy of imaging evidence.The traditional forensics models generally have no thorough functions.When facing the large amount of images from digital devices or network,it is very challenging to search key evidence.In the scenerio of realistic forensic,the proposed model can compelet thorough and standardize functions,each technology being involved in is improved to be more effective and robust.(2)In order to improve the efficiency and accuracy of digital image forensic,this paper studies and proposes the method of image forensic recognition based on multiset feature fusion,select four kinds of imaging features including color,shape,textual and deep feature,moreover,improve the feature fusion method.Firstly,on the basic of traditional CCA(Canonical correlation analysis),a feature fusion method based on D-MCA(Discriminant-minimum correlation analysis)is proposed.MCA removes noise and redundant information from original features by minimum the correlation between features of different modalities,which improves the represent ability of features;combining discriminant information with MCA can make the fused feature achieve more precision performance.To gain more comprehensive features of digital images,secondly extract color moments,SURF(Speeded up robust feature)features,LBP(Local binary pattern)features based on Bo F(bag of feature)model,and extract the deep features using convolutional neural network.At last,the different features fused according to D-MCA are utilizing for implementing forensic recognition of images based on multis-modalities feature fusion.The experimental results on Caltech 101 image set reach an average accuracy of 99.4%,the precision has been improved 49.7% and the recall has been improved 17.5% than single features;the CPU time has been reduced to 11.9% of the traditional fused feature.(3)In order to ensure the authenticity of digital image evidence,this paper studies and proposes a copy-move forgery detection method based match and propagation of sparse-domain.Copy-move forgery detection methods based on dense-domain matching are sophisticated,while the detection methods based on sparse-domain matching are efficient but the detection rate is low.To fill this gap,a propagation mechanism is designed firstly.According to the principle of sparse-domain matching,the SURF algorithm is selected to detect and match the key points of the images,and then the reliable matching pairs are selected as the propagation source by furtherly matching the DWT(Discrete wavelet transform)features of patches.The propagation mechanism is designed to detect more tampered areas by computing the offset(including scale and rotation transformation)between sources,and propagating the matching relationship in the neighborhood of sources.Experiments show that this method achieves detecting rate of 94.0%.The influence of feature selection,rescale transformation and rotation transformation of copy regions on detection rate is analyzed through detailed experiments,which proves that SURF algorithm has more stable performance than others in the matching stage;DWT feature also achieves good performance.(4)In order to ensure the anthenticity and the relevance of digital image evidence,a method of image source identification based on convolution neural network is proposed.A network model with three convolution layers is designed and established for image source identification,which includes three convolutional layers,three regularization layers,three max-pooling layers,one random inactivation layer(Drop-out layer),two fully connected layers and a softmax classifier.The images are divided into 36x36 image blocks using Dresden Image Database 101 data set as experimental data.After the training finished,the test accuracy of the network model reaches 99.8%.Finally,the local-to-global voting strategy is used to get the final decision according to the result of the image block detection.This method can solve the problems of high computational complexity and easily disturbed in noise estimation in traditional methods based on camera characteristics.The convolution neural network in experiment achieves an accuracy of 99.8%,which is 2.6% superior to other similar models.Experimental comparison shows that it has higher accuracy while the time performance is also superior to the traditional methods.