HiDRNet: Hiding image with Deep Reversible Network
Yunlong Li, Weiwei Zeng, Lei Li · 2024
Image hiding task is an important task in the field of information security, where the main objectives are to ensure image quality, payload, and undetectability while hiding secret images. Nowadays, deep learning-based image steganography methods have surpassed most traditional methods, and among these deep learning-based methods, the utilization of a reversible network for image hiding has shown excellent performance. A reversible network is a type of neural network with special properties that enable bidirectional mapping between input and output. It possesses unique advantages in image hiding tasks, ensuring the integrity of hidden information while simultaneously preserving image quality. However, existing reversible neural networks used for image hiding still face security concerns. In this paper, we propose an image hiding method based on a reversible and a scoring network that uses gradient computation to improve the network training process and to achieve higher embedding capacity, lower distortion, and higher security. Further, in the stage of image hiding and recovery, we use dynamic convolution blocks to improve the visual quality to a certain extent while the recovery accuracy is improved. We conducted experiments on the ImageNet, COCO, and DIV2K datasets, and the results indicate that our method improves both image quality and security in terms of comprehensive evaluation metrics, outperforming existing network architectures.