Research On Image Steganography Technology Based On Encoder-Decoder

With the rise of artificial intelligence and Internet of Things technology,people's demands for protecting private information have become stronger and stronger.Information security has become an issue of increasing concern.As one of the important means to ensure information security,information hiding refers to a secure communication method in which secret information is embedded in the multimedia communication cover and transmitted on the public channel.Image steganography is an important branch of information hiding.It hides secret information in public images,which realizes the covert transmission of secret information and guarantees the confidentiality and security of important data effectively.Therefore,it is of practical significance to study the research on image steganography.However,there are still some problems in the current image steganography field.On the one hand,the image steganographic algorithm based on cover generation can realize information hiding in the image.However,it is limited by the data distribution of the training datasets and there are problems such as cover image's quality distortion and insufficient texture complexity.On the other hand,when the existing image steganographic models suffer geometric attacks such as data compression and noise interference during image transmission,it is easy to cause the problem that the secret information is difficult to be completely extracted.In response to the above problems,this paper proposes two image steganographic algorithms based on deep learning:(1)A high-quality image steganography scheme based on conditional generative adversarial networks.We propose a neural network framework with content-aware cover synthesis based on c GAN(CSc GAN),which could generate the suitable cover to embed the secret information for image steganography.This paper introduces semantic label map in the steganography domain,and explores the possibility of steganography by underlying semantic editing.First of all,in order to find suitable steganographic regions in the image,this paper is improving c GAN to find suitable steganographic candidate regions in the semantic label map.Then,we have designed a multi-discriminator c GAN,which can generate high visual quality RGB images from semantic label maps.Finally,our proposed framework is combined with traditional steganographic algorithms(such as LSBM,HUGO,etc.)to embed secret information into the generated steganographic candidate regions.Experimental results show that our proposed algorithm has high imperceptibility and security.(2)A highly robust image steganography scheme based on the Encoder-Decoder framework.The encoder network is used to hide a colorful secret image into another color cover image,and the corresponding decoder is used to extract the colorful secret image.In order to enhance the robustness of the steganography scheme,this paper designs a JPEG compression simulation layer in the neural network to imitate the JPEG compression operation,and enhances the ability of the decoding network to extract secret images by the means of resistant training.In order to improve the ability to extract secret images,the loss function is redesigned to measure the distance between the cover image and the secret image,the secret image and the reconstructed secret image.Besides,steganalysis is introduced to the loss function in order to improve the security of the steganographic algorithm.The experimental results show that our proposed scheme has high robustness.