Research On Image Steganography Without Embedding Based On Generative Adversarial Networks

Image steganography is a technique to hide secret information into carrier images.Traditional embedding-based image steganography is difficult to resist the detection of the increasingly powerful steganalysis.Therefore,researchers proposed image SWE(Steganography without Embedding),which hides information into carrier images without embedding and modifying.With the rise of deep learning,it has been introduced in steganography,especially GAN(Generative Adversarial Networks),which improves the performance of steganography effectively.However,state-of-the-art GAN-based SWEs have some defects in information recovery accuracy,steganography capacity and security model attacks.In view of the above problems,this dissertation designs new SWEs based on GAN.The main work and research results are as follows:Firstly,this dissertation designs a SWE based on adversarial training.The stego images are generated directly from the noise vector mapped by the secret information,then the noise vector is recovered from these stego images by the extraction model.The dissertation designs the generation model with a reasonable feature mapping area,which is conducive to information recovery under the premise of ensuring the quality of the generated images.The extraction model is optimized by widening and deepening the network layers through the residual structure and parallel structure.The wasserstein distance is introduced into the discriminator model,which makes the generated images more natural and improves the model stability.During the training process,we designed a training mode combining stage training and end-to-end training.The training mode ensures different training priorities in different stages to improve steganographic performance.Secondly,this dissertation designs a secure GAN-based SWE model based on modern cryptography security principles.The security steganography model based on asymmetric keys is designed for the known generation model attack,the known extraction model attack and the training extraction model attack.The generation model and the extraction model are set keys separately.The sender generates stego images by the encoding key and secret information.After the decoding key and the stego images are obtained by the receiver through the secure channels and the insecure channels respectively,the secret information is recovered by the extraction model.The model can complete the secure transmission of secret information under the premise of correct encoding key and decoding key.Finally,this dissertation proves the effectiveness of the model from different perspectives through experiments.The experimental results show that the model can recover information with high accuracy,improve the steganography capacity greatly,and have good security.The secure GAN-based SWE model can achieve the secure transmission of secret information in insecure channels on the premise of ensuring high recovery accuracy.The model can resist various security attacks,and meet the Kerchhoffs principle of modern cryptography.