Research On Digital Image Operation Forensics Based On Deep Learning

In recent years,with the development of image processing technology,it is becoming more and more easy for people to edit images.When people enjoy the convenience brought by the image technology,they also have the corresponding problems,that is,the authenticity of the image has been challenged.The development of Internet makes our lives full of image information.However,malicious tampered pictures are not easy to identify.Once the false images are disseminated,they will have a great impact on military,politics,judicial,news and other fields.Therefore,it is an urgent problem to be solved to be able to detect the authenticity of digital images.In the process of image editing,different image operations will be used to change the content of the image.If you can detect whether the image has undergone some operations,it is of great significance for identifying the original and authenticity of the image.In this paper,based on the intensive study of the detection algorithms of digital image operation at home and abroad,a solution for digital image operation forensics based on deep learning is proposed.The main work of this article is as follows:(1)Existing median filtering detection methods are no longer effective for small size or highly compressed images.To deal with this problem,a new median filtering detection method based on CNN is proposed in this paper.Specifically,a new network structure called MFNet is constructed.First,for preprocessing,the nearest neighbor interpolation method is utilized to up-sample the small-size images.The property of median filtering can be well preserved by the up-sampling operation and enlarged difference between the original image and its median filtered version can be obtained.Then,the well-known mlpconv structure is employed in the first and second layers of MFNet.With mlpconv layers,the nonlinear classification ability of the proposed method can be enhanced.After that,three conventional convolutional layers are utilized to finally derive the feature maps.The experimental results show that the proposed method achieves significant improved detection performance.Moreover,the proposed method performs well for highly compressed image of size as small as 16×16.(2)To detect various image operations,in this paper,we propose a carefully crafted CNN model to learn features from the magnified images and do multi-classification automatically.Firstly,the images will be magnified by nearest neighbor interpolation in the preprocessing layer.Then,hierarchical representations of different operations are learned via multi scale convolutional layers.The mlpconv layers are used to enhance the whole architecture's nonlinear modeling ability.Further more,shortcut connections between mlpconv layers allow for increasing the depth of the network while reducing information loss.We present comprehensive experiments on 6 typical image operations.The results show that the proposed method has a good performance both in binary and multi-class detection effectively.(3)Aiming at the problem of inadequate use of different level features and loss of transmission information in the above CNN models,a multi-level feature fusion structure is designed to detect many kinds of image operations.In order to enhance the information transmission between the bottom and the top of the model,and improve the reusability of the feature,in the proposed network,different levels of feature fusion structure are added between the Mlpconv layers,namely,each layer is connected to the first two layers.The proposed model is applied to detect multi-class image operations,with an average accuracy of over 95%.It provides a reliable solution for multi-class image operations detection.