Research On Passive Forensics Of Digital Image Color Tampering Based On Deep Learning

As one of the most important information carriers,digital image plays a vital role in people's daily life,entertainment and work.With the widespread spread of digital image,the need to modify digital images has become more and more urgent,resulting in various image editing software such as Photoshop,Meitu Xiuxiu and so on.Although these image editing software have brought great convenience to people,they have also given the opportunity for lawbreaker.If these fake images are used in the fields of media,scientific research,and justice,they will inevitably cause adverse effects.Therefore,digital image forensics has caused widespread concern among researchers.Image colorization technique,that is,image coloring,is a technique that colors a grayscale image by a certain algorithm and causes the human eye to not perceive changes.Although image coloring is one of the most common image operations in image editing,forensic research on image color tampering detection has only begun to attract researchers' attention in the past two years.This paper has carried out in-depth research on digital image color tampering passive forensics.A three-phase fake colorized image detection method based on image channel statistical characteristics and neural network is proposed.First,the method compares the statistical differences between the color distribution of natural colorized images and the color distribution of fake colorized images.Thereafter,feature extraction is performed using statistical differences in the presence of digital images.Finally,the extracted feature data is used to train the well-designed neural network to complete the detection of the authenticity of the colorized image.An end-to-end fake colorized image detection method based on multi-model fusion is proposed.The method constructs a multi-model fusion double-flow convolutional neural network structure consisting of a classification sub-network module,a feature extraction sub-network module and a fusion output module.The ability to extract the color features of the digital image is improved by integrating the classification information output by the classification sub-network module and the feature information output by the feature extraction sub-network module.Experiments show that the two methods proposed in this paper can effectively identify the authenticity of color images,and also show good performance advantages compared with the existing detection methods.