CAUNet: Cross-Attention Unet for Image Steganography

Image steganography is an important branch in the field of information hiding, aiming to covertly embed secret information into the cover image. Existing UNet-based steganographic methods demonstrate strong feature extraction capability, but the simple skip connection performed between the encoder and decoder often leads to low feature fusion efficiency and easily introduces redundant information. To solve this problem, this paper proposes a new network structure called CAUNet (Cross-Attention UNet) for image steganography. The core contribution lies in designing a lightweight cross-attention module and embedding it into the skip connection of UNet. This module dynamically learns the interdependence between encoder and decoder features to generate an attention map for weighting and refining the encoder features, thereby achieving more effective and targeted feature fusion. Experimental results show that, compared with the conventional UNet-based steganography model, CAUNet significantly improves the visual quality of the reconstructed stego images, demonstrating that the proposed cross-attention mechanism effectively enhances the performance of steganographic networks.