Research On Key Technologies Of Digital Image Forensics Based On Deep Neural Network

With the development of image editing software,the cost of image editing is further reduced,and the malicious editing of images has caused serious problems to people's lives.At the same time,with the advent of generative adversarial networks(GANs),it has become easier to generate high-quality fake images.However,there is currently no forensic algorithm that can detect both artificially tampered images and GANs-generated images at the same time,and the forensic algorithms is not strong in generalizing the images generated by GANs.Digital image forensics is an issue worthy of further exploration.In this paper,we discuss some key issues in the field of image forensics and the deep learning techniques.we propose a digital image forensics method based on deep neural network.The method proposed in this paper can detect tampered images and GANs-generated images simultaneously,and the detection Macro-F1 score achieved 0.9865.In addition,compared with previous work,this method achieves better performance in the task of detecting only tampered images,and has strong generalization performance for images generated by various GANs models.The main work of this paper is as following:1.To solve the problem that current digital image forensics methods cannot detect tampered images and GANs-generated images simultaneously,we propose to take the edge of the object in the image as the unified detectable feature.The Scharr operator is selected as the edge detection operator according to its accurate edge detection,rich edge details and distinct difference between the edge of the edited area and the edge of the real area in the fake images.2.To solve the problem that it is difficult to detect forgery marks in tampered images and GANs generated images,we propose that two-phase subtraction of R,G,B grayscale channels in RGB color space can suppress the details information of the objects in an image,and highlight the edge information of the tampered area.We provide a general formula that can be optimized by adjusting the parameters,and also explain the reason that why images are usually converted from RGB color space to YCr Cb color space in the previous digital image tampering forensics tasks.3.The size and the format of the digital image to be detected are uncertain,while the deep neural network usually requires the input data have the same size in a mini-batch.The traditional image scaling method will change the details of the fake image to some extent and interfering with the detection of forgery marks,so they are not suitable for the digital image forensic task.To solve this problem,we propose that the gray level co-occurrence matrix(GLCM)can be used to normalize the images which as the input of the deep neural network.4.According to the characteristics of the depthwise separable convolution with few parameters and can decouple channel dependence and spatial dependence of the feature map,a deep neural network based on depthwise separable convolution is designed,which results in fewer network parameters so that improves the detection efficiency.It shoud be noted that the detection performance of this network model is not reduced compare with the traditional convolutional neural network.