Research On Image Steganography Based On Deep Learning

The rapid development of Internet produces quantities of data in the network.The private data are easy to be attacked in transfer process.Steganography can protect the private data by hiding it into multimedia data.Traditional steganographic algorithms have high requirements for designers in terms of minimum distortion function and steganography coding mode.The steganographic model based on generative adversarial network utilizes the game process of steganography and steganalysis to naturally embed information into the image.However,current mainstream steganographic models have some shortcomings in embedding capacity,imperceptibility and robustness,so this paper improves the overall steganographic performance of the model by designing and optimizing the network structure.The main contributions of this paper include:Firstly,a steganographic model based on residual and dense networks is proposed to increase the embedding capacity.The encoder,decoder and critic in the model form an endto-end steganographic framework.The encoder and decoder realize feature reuse and improve the training efficiency and stability.The critic scores the cover image and the stego image,and uses Wasserstein distance as the measurement index for training,so as to ensure the stability of training and make the generated stego image of higher quality.The model can not only embed large amount of information but also ensure the quality of the image and the accuracy of decoding.Secondly,a robust steganographic model based on generative adversarial networks is proposed,which combines modules containing different kinds of noises with the steganography model to train a robust steganographic model that can resist various attacks.Image noise mainly includes clipping,scaling,JPEG compression,etc.,which is packaged into modules to be added between encoders and decoders.The encoded image generated by the encoder is processed by different kinds of noise,and then decoded by the decoder,which ensures that the encoded image can have a higher decoding accuracy even in the face of various noise attacks in the real environment.Finally,this thesis analyzes the proposed steganographic model in contrast with the mainstream steganographic model,and uses various indicators for quantitative analysis.The image quality is mainly reflected in the PSNR and SSIM indexes,the embedding capacity is measured by Bpp,and the robustness is reflected by the decoding accuracy and the detection accuracy of the steganalysis model.In the contrast analysis of noise layer,JPEG compression adopts a more reasonable differentiable simulation method,which makes the encoded image still have a higher decoding accuracy in the real JPEG compression scene.The results show that the model proposed has a significant improvement of performance compared with other steganographic models.