Research On Generative Image Steganography Based On Deep Learning

Steganography is a communication technique or storage method based on information hiding,in which secret information is hidden in a carrier that can be made public through an imperceptible method.During the long confrontation development of steganography and steganalysis,researchers have proposed a wealth of ideas and methods for image steganography.The non-embedded image steganography algorithm does not modify carrier image,although theoretically it has high security,but generally it has the problem of single image characteristics and poor quality of the generated image.Embedded image steganography algorithm writes secret information by modifying the carrier image,so it is easy to leave modification traces and be attacked by steganalysis models.To deal with the problems of existing steganography methods,this dissertation proposes different image steganography algorithms in a targeted way.Firstly,this dissertation proposes a generative steganography algorithm based on the self-attentive mechanism from the embedding-free steganography approach.With the sample generation capability of generative adversarial networks,the secret information is used to generate the secret-laden images directly.Also,based on the existing work with the problems of poor image quality and image background anomalies,the self-attention mechanism is designed and implemented to correct the distortion and artifacts of image generation by the calculation of global pixel correlation.The steganograph discriminator loss function is designed to fit the soft interval loss,which greatly improves the quality of the generated images on the basis of guaranteeing the capacity of secret information.The extractor resolves the problem of image quality degradation caused by overoptimization of parameters through the setting of accuracy threshold.Secondly,this dissertation proposes an image stylized steganography model based on self-coding network from embedded steganography.According to the feature that the image adaptive regularization can achieve arbitrary style migration,a network model that can perform steganography in the process of image stylization is designed and implemented.The embedded steganography security is improved to a certain extent by covering the embedding of secret information through the act of active artistic processing of images.For the image stylization process,the image content will be substantially preserved phenomenon,the learnable preprocessing filter is designed to extract the image texture content features as the embedding probability and hide the information in the image content.The receiver filters the stylized image and recover secret information.Finally,experiments are designed to demonstrate the effectiveness of the proposed algorithm model,and the performance of the algorithm is measured by different evaluation metrics.The experimental results show that the proposed algorithm in this dissertation can generate higher quality near-natural images;the designed model is able to perform embedding and extraction of secret information,and the capacity is also better than existing steganography methods.