Research On Steganographic Distortion Function Learning Method For Color Image Based On GAN

Adaptive image steganography has been becoming a hot topic,as it conceals covert information within the texture region of an image by employing a defined distortion function,which guarantees remarkable security.In spatial gray-scale image steganography,the research on automatically generating steganographic distortion using the generative adversarial network has achieved a significant breakthrough recently.However,to the best of our knowledge,there are not related works in spatial color image steganography.Compared with the gray-scale image,color image steganography should preserve the channel correlation and reasonably assign the embedding capacity among RGB channels simultaneously.In this paper,adaptive steganography of color images is studied,and two methods for designing spatial color image steganography distortion function based on generative adversarial network are proposed.(1)In this paper,a new structure of generating adversarial network for designing color image distortion function is proposed.In the generator network,the spatial dependence of the adjacent pixels and channel dependence of the color image are carefully considered,and a parallel network structure is designed to obtain the feature maps containing rich spatial and channel related information,and the generated embedded distortion function can effectively reduce the destruction of the correlation between the three channels of the color image.For the color image carrier,the gray image steganalysis is modified as the antagonistic part of the network.Also,the generator can automatically learn to allocate the embedding capacity for three channels via controlling the total steganographic capacity in generator's loss function and alternately training the discriminator.The experimental results show that our proposed framework outperforms the advanced spatial color image steganographic schemes in resisting the color image steganalysis.(2)In this paper,a twin network structure is designed as a generator to generate the steganographic distortion function of color images,and a new subnetwork is used to learn the adjustment of modification probability.The first subnetwork of the generator generates the modification probability matrix,and the second subnetwork adjusts the positive and negative modification probabilities.The positive and negative modification probabilities are reasonably allocated through network learning.In the loss function of the generator,the total embedding capacity of the three channels of the color image is also controlled so that the network can reasonably allocate the embedding capacity of the three channels by learning.The experiment proves that the designed twin network structure can further reduce the destruction of color image channel correlation,which fully demonstrates the feasibility of using traditional methods for reference in the deep learning design process.