Low-frequency and Robust Image Information Hiding

zhiyi cao, Jiwei Zhang, Ruiqiang Guo, Zili Zhang, Huo Lina

Keywords: Information hiding, Contrastive Learning, Invertible Neural Network

Abstract: Existing image information hiding methods commonly lack robustness to varying degrees of distortion on container images. This paper proposes a low-frequency and robust image information hiding method, LRIIS, to overcome current challenges. To emphasize robustness, instead of hiding high-frequency subbands, we propose a novel wavelet contrastive loss to constrain so that most secret information is hidden in the low-frequency subbands. Compared to high-frequency subband hiding, low-frequency subband embedding achieves enhanced robustness. To alleviate the varying degrees of distortion influence, we build an unsupervised Attacked Image Enhancement Module (AEM) to generate the de-attacked image that is close to the corresponding container image. Notably, thanks to the pseudo-class label of AEM, the proposed method can recover the secret image from the attacked image without requiring a specific attack label. Experimental results demonstrate the superior performance of the proposed LRIIS model on the COCO and DIV2K datasets compared to existing state-of-the-art image information hiding methods.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

Submission Number: 6462