DASS-Net: Degradation-Adaptive Semi-Symmetric Network for Robust Image Hiding

Previous INN-based image hiding methods have demonstrated outstanding concealing and revealing performances; however, their robustness is compromised due to the unbiased and lossless characteristics inherent in strictly symmetric INNs. To address this limitation, we propose the Degradation-Adaptive Semi-Symmetric Network (DASS-Net) for robust image hiding. DASS-Net incorporates asymmetric units into INN flow, injecting nonlinearity and adaptability to the bijection. Specifically, Coupled Learning Unit (CLU) is proposed as asymmetric unit in the concealing phase, to facilitate more flexible information interaction. While the revealing process incorporates Degradation-Aware Modulation Units (DAMU) as asymmetric units that capture degradation clues and customize modulation parameters, thus enabling dynamic adjustments for optimal revealing under various degradation scenarios. Additionally, we propose a Frequency Sub-bands Enhancement Module (FSEM) that leverages low-frequency features to assist the recovery of high-frequency features and strengthen hidden information, thereby mitigating information loss caused by degradation. Extensive experiments on COCO and DIV2K datasets, exhibiting a range of degradation types and degrees, indicate that DASS-Net markedly surpasses state-of-the-art methods, with more than 10% improvements in imperceptibility and robustness.