Research On Anti-forensics Of Digital Image Contrast Enhancement Based On Generative Adversarial Network

With the development of computer technology,digital images are widely used in people's life.At the same time,the potential security problems in digital images are becoming increasingly serious.The spread of tampered digital images through the Internet and other channels will affect people's judgment of things,and even cause harm to the economic and social order.To detect tampered digital images and protect information security,digital image forensics technology has been greatly developed in recent years.At the same time,to explore the security of forensics technology and perfect forensics theory,digital image anti-forensics research has also received extensive attention.Aiming at digital image contrast enhancement anti-forensics,this paper proposes three anti-forensics algorithms based on Generative Adversarial Network from different angles while maintaining the visual effect brought by image contrast enhancement operation,to further explore the security of existing forensics algorithms.(1)A contrast enhancement anti-forensics algorithm based on Generative Adversarial Network is proposed.The current contrast enhancement anti-forensics algorithm has weak anti-forensics performance against the forensics method in the pixel domain.In order to further explore the security of this kind of forensics algorithm,this paper proposes an anti-forensic framework based on Generative Adversarial Network to achieve contrast enhancement anti-forensics.A generator structure based on multiple residual learning is designed,and a new feature loss term is introduced to train the generator.Experimental results show that the algorithm performs well in anti-forensics attacks,and can maintain the visual effect of contrast enhancement with high image quality.(2)A contrast enhancement anti-forensics algorithm based on Generative Adversarial Network guided by prior knowledge is proposed.The algorithm uses the prior knowledge of contrast-enhanced image histogram traces to creatively construct two loss functions and uses them to guide the training of the network.At the same time,to make sure that the back-propagation of the gradient information guided by the newly designed loss term is valid in the training process,specific mean-shifted Gaussian functions are designed to calculate the derivable image histogram.Experimental results show that the anti-forensics attack performance of the proposed algorithm is better while maintaining high image quality.(3)An anti-forensics algorithm of the “JPEG-contrast enhancement” operation chain is proposed.At present,there are few anti-forensics algorithms of the image operation chain.In order to explore the security of operation chain forensics algorithm,a hierarchical generative adversarial mechanism based on Generative Adversarial Network is designed to process the image of “JPEG-contrast enhancement”.In addition,a loss term based on the relationship between adjacent pixels is introduced to reduce the image quality loss caused by JPEG compression.Experimental results show that the algorithm can achieve better anti-forensics attack performance while maintaining high image quality.