Image Source Forensics Based On Deep Learning

With the development of the information technology and the popularization of digital camera equipment,image has become an important information carrier.At the same time,the widespread use of image editing tools has challenged the authenticity,primitiveness and integrity of images.Image tampering with malicious purposes will bring serious negative effects to society.In order to effectively curb the behavior of image forgery,digital image forensics technology came into being.As an important branch of digital image forensics,image source forensics plays an important role in copyright protection and image traceability.In this paper,the accuracy,security and applicability of the image source forensics are studied and analyzed.The main achievements include:(1)An attack scheme for recaptured images forensics algorithm based on generative adversarial network is proposed.From the attacker's point of view,this paper proposes a recaptured image conversion scheme,aiming at eliminating the texture traces introduced by recapturing and preventing the image from being detected by the classical recaptured forensics scheme.Based on the generative adversarial network,we design the network structure of generator and discriminator according to the characteristics of recaptured images,and creatively proposes a new loss function constraint to avoid large distortion of image content.The experimental results show that the scheme can transform the recaptured images from visual effects and statistical features,achieving double deception of human vision and image source forensics algorithms.(2)A recaptured images forensics algorithm based on deep learning is proposed.From the point of view of algorithm security,this work aims to design a scheme that can effectively resist the attack of adversarial examples.According to the generation principle of adversarial examples,two security improvement strategies are proposed based on the general deep learning framework.in this work.Firstly,a penalty term is added to the loss function to control the smoothness of the derivative of cross-entropy with respect to the input samples.Secondly,the strategy of fusion of normal samples and adversarial examples is adopted to train the network.The experimental results show that the proposed scheme can classify the adversarial examples well while maintaining the high classification accuracy of the normal samples,and the security of the proposed scheme is enhanced.(3)A source camera identification scheme for recaptured images is proposed.In this work,the camera source forensics technology is applied to the recaptured images,which enlarges the applicability of the camera source forensics technology.Because the recapturing will introduce a lot of noise to the image and interfere with the original features related to the camera,the common Source camera identification algorithm is no longer fully applicable.In order to extract more abundant camera features,the scheme adopts the network structure of sub-network parallel connection.Different sub-networks adopt different high-pass filters for preprocessing.And the features extracted from sub-networks are combined together for classification module.The experimental results show that this scheme can effectively improve the accuracy of camera source forensics for recaptured images.