Research On Source Forensics Of Digital Images And Videos

We are currently on a new digital era and an information age when image,video,audio,text,etc,interacts.With the rapid development of imaging technology and gradual maturity in product manufacture,there is a population explosion among the devices of digital media acquisition,such as digital camera,smart-phone,and so on.Meanwhile,development in artificial intelligence has accelerated digital media editing techniques that give impressive visual effect.Everyone can easily manipulate a digital media via various editing tools,and then share it with the others after uploading to social network platforms.It is a great challenge for the verification of digital media authenticity and integrity.Therefore,studying on new technique for multimedia forensics is an imminent task.Targeting at the digital image and video,this thesis make the analysis of source forensics in three aspects: capturing scenes,capturing devices,and editing software.The main contributions are as follows,1.we propose a generalized scheme based on Laplacian Convolutional Neural Networks for recaptured image forensic.To the best of our knowledge,it is the first deep learning-based work for recapture forensics.We embed a Laplacian filter into Convolutional Neural Networks structure,which attenuates the disturbance of image-contents and strengthens the difference between the original and recaptured images.We evaluate the proposed method on four kinds of small-size image databases.The experimental results have demonstrate that the proposed algorithm is effective.2.we propose a deep learning-based solution: content-adaptive fusion residual networks,to source camera identification.Firstly,we explore self-learning convolution as preprocessing to amplify forensic-related signal and design a multi-scale fusion residual network to capture more comprehensive features.Then,according to the differences of image contents,the images are divided into three subsets and train three fusion residual networks through transform learning technique.The experiment results show that the proposed method has satisfactory performances at three levels of source camera identification: brand level,model level,and device level.Comparing with PRNU-based method,proposed method achieved better performance on lower-resolution images.3.we come up with a new proposal about device forensics of high dynamic range image and collect a novel dataset of High Dynamic Range(HDR)and Standard Dynamic Range(SDR)images captured by smartphones is presented.Then,we analysis the effect of HDR image on sensor noise-based source identification algorithms and find the problem of pixel shifting during aligning multi-exposure images.Based on the above observation,we propose a block-consistency feature for camera identification of HDR images.Specifically,in order to keep the variousness of the dataset,the images are acquired in different conditions,including various capturing motions,scenes and devices and there are 5415 images which consists of HDR and SDR image pairs.It is the first public dataset of HDR images for image forensics.In addition,proposed dataset has been exploited to evaluate the performance of sensor noise-based source identification algorithms when applied to both kinds of images.Result have proved the robustness of this method,but also the difficulties in source identification of HDR images.Focusing on source camera identification of HDR image,we propose a block-consistency feature.Experimental results demonstrate its effectiveness.4.we introduce a container-based method for the analysis of video integrity.Especially,the video containers are characterised by the set of symbols(fieldsymbols and value-symbols)and converted into a vector.Then a decision tree based classifier is applied to a vectorial representation of the video container structure and enriched with a likelihood ratio framework designed to automatically clean up the container elements that only contribute to source intra-variability.We conducted an extensive validation on a new build dataset of 7000 video files and the proposed method achieves high accuracy in distinguishing pristine from tampered videos and classifying the editing software,even when the video is cut without re-encoding or when it is downscaled to the size of a thumbnail.In addition,the proposed method is both efficient and effective and can also provide a simple explanation for its decisions.