RSAINN: Residual-Spatial Attention Based Invertible Neural Network for Robust Image Steganography

Image steganography enables covert communication by embedding secret information into cover images in an imperceptible manner. However, existing approaches still encounter significant challenges in maintaining imperceptibility and ensuring security, especially when handling images with complex textures. To overcome these limitations, this paper proposes a novel steganographic network based on Residual Spatial Attention Invertible Neural Network, termed RSAINN. The network leverages its strong ability to extract features, effectively capturing texture and boundary information, and employs an attention mechanism to guide feature learning by selectively focusing on key regions. The proposed Residual Spatial Attention (RSA) module integrates a multi-level skip connection mechanism to capture rich local features. This combined attention mechanism effectively mitigates the degradation in visual quality and the weakening of security performance in stego images. In addition, we optimize the network architecture by redesigning the convolutional blocks and increasing the network depth, thereby enhancing the ability to extract fine-grained details. Extensive experiments conducted on the DIV2K, COCO, and ImageNet datasets demonstrate that the proposed method leads to a significant improvement in the visual quality of stego images, achieving PSNR gains of 1.8 dB, 10.48 dB, and 7.52 dB, respectively, compared with baseline methods.