Research On Texture-aware And Scale-varying Deep Image Steganography Algorithm

Image steganography plays an important role in information hiding and privacy protection,which achieves secure transmission in public channels by embedding impor-tant information into images,such as digital certificates and secret keys.Deep learning,an emerging technique to implement steganography,makes up for the shortcomings of traditional methods that require hand-designed features.However,there are still some fatal deficiencies in current deep image steganography models.They don't consider the texture complexity difference between the pixels of the carrier image at different posi-tions,so that the stego image suffers from color distortion and brightness distortion in visual quality.They can only deal with the secret information with fixed size,and ig-nore the loss caused by data type conversion or JPEG compression,which greatly limits their practicability.To address these problems,we conduct the following studies:1.A scale-varying deep steganography analysis model(RMCNN)is proposed.This model is a 50-layer convolutional neural network with residual blocks,which im-proves the accuracy of steganalysis by optimizing residual calculation,feature extrac-tion and classifier at the same time.RMCNN introduces a moment layer that extracts a fixed-size statistical moment feature with the global pooling,providing the necessary support for steganography.2.A texture-aware scale-varying deep image steganography model(STGAN)is proposed,which can hide scale-varying secret information(including images,text,and binary)into natural images.This model is an end-to-end network,including hiding network,reveal network and steganalyzer,while the first two subnetworks are fully convolutional networks that keep the size of feature maps unchanged,and the stegan-alyzer is RMCNN,so that STGAN can detect images of different sizes.In order to improve the visual quality of stego images,a loss function TL based on texture com-plexity is proposed to give each pixel a penalty weight that is inversely proportional to complexity of image texture,embed more information in rich texture areas.To enhance the robustness of the hiding network,we construct a noise layer including a truncation layer and a JEPG layer to ensure that it can recover the secret information from the stego image with corresponding loss.In addition,when the secret information is text or binary,we can adopt QR code as an intermediate carrier without changing the structure of STGAN to achieve scale-varying information steganography.3.We conduct abundant experiments to validate the effectiveness of STGAN.The STGAN trained has strong robustness.It can take any natural image as a carrier,and scale-varying images or text as secret information to generate stego images with excellent visual quality.The average SSIM between cover and stego images is greater than 0.98.When the cover image is in Bossbass database and the embedding capacity is 1 bpp,the accuracy detected by ATS is 0.523,indicating that STGAN guarantees the security of secret information.