ROSIN: Robust Semantic Image Hiding Network

Yuan Zhao, Bo Liu, Tianqing Zhu, Ming Ding, Wanlei Zhou

Abstract: Image steganography aims to imperceptibly conceal a secret image within a cover image, allowing for its recovery at the intended receiver without arousing suspicion. However, many current approaches prioritize high output quality and payload capacity, often neglecting robustness, especially when images are subjected to distortions during transmission. To address this limitation, we propose ROSIN (Robust Semantic Image Hiding Network), a novel framework that enhances the robustness of hidden images by leveraging the stability of semantic features in cover images. Specifically, our method semantically disentangles the cover image into identity and attribute features, and then embeds the secret image into the identity feature, which is more resilient to distortions due to identity feature’s value and geometric invariance properties. We evaluate ROSIN across multiple dimensions, including imperceptibility, fidelity, and robustness, using metrics such as PSNR and SSIM. Experimental results demonstrate that ROSIN achieves superior robustness, outperforming SOTA methods by approximately 15% in resisting distortions, such as JPEG compression and noise attacks, while maintaining comparable imperceptibility and fidelity with PSNR of 44.97 dB and an SSIM of 0.97 for 256x256 secret image. Our method strikes a balance between imperceptibility and robustness, making it suitable for real-world applications where image transmission is subject to various lossy distortions. This work also represents the first attempt to leverage the stability of semantic features to enhance the robustness of image steganography, opening a new research direction.