Forensic Algorithms For Digital Image Manipulations

With the wide application of camera equipment,digital images as an important information carrier,is gradually integrated into our life.As softwares for image editings become more and more popular,people gradually realize that seeing by eyes may not be reliable.In recent years,the frequent occurrence of image forgery events in the fields of news media,judicial notarization and scientific research makes it an urgent problem to distinguish the authenticity and originality of digital images during evidence collection.Tampers often change the image and its related information through different image operations,which affects the readers' judgment on the authenticity.Therefore,it is of great significance to detect the operation history on digital images.Based on the above research background,this paper first summarizes the research status of image operation forensics algorithm,and then proposes effective solutions from the accuracy,practicability and robustness of forencis algorithms for digital image manipulations,mainly including the following work:(1)A forensic algorithm of digital image manipulations for small-scale images based on the difference enhancement between classes has been proposed.In order to improve the performance of detection algorithms on small-scale images,this work designs algorithm from two levels of preprocessing layer and classification network,which can detect six typical image operations.At first,the preprocessing layer enhances the inter class discrimination without introducing additional information,so as to prepare for the next step of feature extraction and classification;at the same time,in order to adapt to the problem of image size reduction and image operation trace reduction,the Xception network structure is introduced to optimize feature extraction.Results in this section show that this method can accurately detect many kinds of image operations,even in more challenging scenarios.(2)A forensic algorithm of digital image manipulations based on fusion preprocessing has been proposed.We only take accounts the difference between enhancement classes and the inhibition of image content is not sufficient in the preprocessing process.In the first section,I further optimizes the preprocessing layer in the following ways.On one hand,the nearest neighbor interpolation algorithm with scale two is used to increase the inter class differentiation;on the other hand,the high pass filter preprocessing with fixed parameters is integrated into the end-to-end CNNnetwork to improve the suppression on image content.In the second part of my work,the ordinary convolution layers are replaced by the depthwise separable convolutions layers to improve the performance of the model without increasing the complexity of the algorithm,too.The studies show that the accuracy of this scheme is improved in the detection of small-scale images with not only multiple single-operation types and but also multiple operation chain types.(3)A forensic algorithm of digital image manipulations with JPEG robustness has been proposed.The existing multi-operation forensics schemes are not robust to JPEG post-processing,often with poor detection performance.This algorithm includes two parts: preprocessing layer and classification network,which can effectively extract the image operation traces weakened by JPEG post-processing,improves the JPEG robustness of the algorithm.First,using the self-learning preprocessing convolution layer instead of the fixed-parameters filter kernel set by experiences can effectively suppress image content and make the network more focused on learning and extracting weak image operation trace features.Second,in the classification network,the downsampling module and feature extraction module combine depthwith separable convolution and residual connection,so as to improve the non-linear expression ability of the network and ensure the ability to learn weak tracing features.My work shows that this method performs well in all JPEG robustness detections with both high and low quality factors.