| Information steganography as an important method of information hiding,which mainly focuses on embedding information into an image in a secretive manner without changing the characteristics of the image,has attracted a lot of attention in the field of information security.Traditional image steganography relies on the in-depth a priori knowledge of researchers when designing the algorithms,thus posing many difficulties in algorithm design as well as applications.The emergence of deep learning has broken this barrier,and researchers have combined deep learning with image steganography to drive the rapid development of image steganography research.However,existing image steganography still has certain problems in terms of quality,invisibility,information extraction accuracy,and robustness of generated images.In addition,the detection performance of steganalysis has been continuously improved,threatening the security of the model tremendously.Thus,we conducted research on image steganography based on GAN and attention mechanisms in this paper.The main research is summarized as follows:(1)To address the problems of low quality of generated steganographic images,large discrepancies between revealed secret images and original secret images,and the limited capability of anti-steganalysis detection,we propose an image steganography framework with multi-scale features integration based on GAN called MFISGAN.To begin with,we constructed the secret image feature extraction network(Sfe Net)and the encoder to achieve the extraction and integration of multi-scale features of the secret image with the feature information of the cover image.Additionally,the steganalyzer is introduced as the discriminator.The WGAN-GP optimized objective function is used to alternately update the parameters of the encoder and the discriminator to simulate the adversarial to enhance the model to resist steganalysis.Finally,a novel mixed loss function is proposed by combining perceptual loss,MS-SSIM,and L1 loss.Experimental results show that,in comparison with existing methods,MFISGAN effectively improves the image quality and the security of the model.The average PSNR and SSIM of steganographic images reached 42.95 d B and 0.9887,respectively.The accuracy of steganalysis detection decreased by more than 20%.(2)To address the problems of insufficient noise resistance and low accuracy of labeled information extraction for the steganographic images generated by the existing models,an attention mechanism-based anti-noise robust image steganography model-AN-RISNet is proposed.Various basic modules incorporating attention mechanisms and residual networks are designed for construction of the preprocessing network,encoder,and decoder.Specifically,the preprocessing network performs channel expansion and size transformation of the labeled image to increase the redundancy space of the labeled information.And the encoder and decoder are applied to the embedding and extraction of labeled information.Moreover,the simulated noise layer is added between the encoder and the decoder to make the generated steganographic images resistant to multiple intensities and types of noise attacks by using the combination of encoder-noise-decoder training.Finally,the binary cross-entropy is introduced in the loss function to measure the loss before and after the embedding of the labeled information.Experimental results show that,in the face of different types and intensities of noise attacks,AN-RISNet generates steganographic images with an average PSNR and SSIM exceeding 40 d B and 0.98,respectively.And the average extraction accuracy of labeled information is more than 90%.(3)The image steganography system is designed and implemented according to the research content of this paper.The system consists of two modules: secretive communication and copyright protection.The MFISGAN and AN-RISNet are also applied to the system to achieve embedding and extraction of information. |
