Research On Deep Learning-based Image Steganography With High Security

Image steganography aims to hide secret information into public images,and transmit it to the receiver through public channels without attracting the attention of attackers.The traditional image steganography algorithm adopts the same embedding strategy for any image,and the steganalyzer can detect the stego image by analyzing the statistical features of the image.The content-adaptive image steganography assigns appropriate embedding distortion to each pixel of the cover image by defining a distortion function,and then embeds information by minimizing the total distortion for a given embedding capacity.The content-adaptive steganography algorithm greatly improves the security of steganography,but the definition of the distortion function relies on manual design,which requires researchers to have rich experience and spend a lot of time to adjust the parameters of the algorithm.Some researchers have proposed a steganography model based on CNN encoder-decoder networks,which can greatly improve the capacity of steganography,but the generated stego images have serious distortion and low security.Moreover,with the emergence of steganalyzers based on convolutional neural networks,the detection performance of the steganalyzer has been greatly improved.In order to improve the security of steganography algorithm,this paper uses generative adversarial network and adversarial examples to improve the anti-steganalysis ability of steganographic algorithm,which are as follows:1)Aiming at the low security of steganography methods based on encoder-decoder networks,an encoder network is designed to better fuse the cover image and the secret image,and the generative adversarial network is used to improve the anti-steganalysis ability of the secret image.In this paper,Small-Inception blocks are used to construct an encoder network to better integrate the features of different receptive fields.At the same time,encoder network is trained against the steganalyzer,so that the encoder network learns to embed information in the position that the steganalyzer cannot detect,and the generated stego image has better anti-steganalysis ability.The experimental results show that,compared with Atique’s steganographic method,the stego image generated by the proposed method is more visually similar to the original cover image,and the PSNR and SSIM between the cover image and the stego image are 34.89 and0.96,respectively.At the same time,our method has better security,with a detection accuracy of 78.9% on Xu-Net.2)Aiming at the problem that adversarial embedding methods need to design perturbation for each input image and the transferability is poor,a generative network is designed to generate universal adversarial perturbations to improve the security of content-adaptive image steganography method.In this paper,the generated network will learn the entire data distribution of common disturbances by training it to attack steganalyzer.The generated perturbations are added to the modification probability of the image to change the embedded location of part of the information that will send the stego image to the decision boundary of the steganalyzer,causing the steganalyzer to misclassify.The experimental results show that the security performance of the content-adaptive steganography methods after adding perturbation is improved by10.84%,and the security performance on the non-target steganalyzer is improved by8.15%.Even if the steganalyzer is retrained with stego images generated by our method,a performance improvement of up to 2.78% can still be achieved.