Robust Invertible Image Steganography for Noisy Channel Transmission

초록

Invertible Neural Networks (INNs) have emerged as a powerful paradigm for high-capacity image steganography due to their bijective mapping properties and theoretically perfect reconstruction in ideal environments. However, INNs exhibit severe vulnerability to spatial-domain perturbations, where even minor Gaussian noise can be heavily amplified during the inverse transform, leading to cross-channel artifacts and catastrophic extraction failures in noisy communication channels. To address this challenge, we propose a Stego Denoising Module (SDM), designed as a lightweight residual convolutional network that acts as a plug-in feature rectifier for the INN-based extraction pipeline. Furthermore, we introduce an SNR-adaptive gating mechanism that selectively activates the SDM based on real-time channel conditions, preventing unnecessary 'over-processing' distortions under favorable communication environments. Experimental results demonstrate that the SDM effectively suppresses structural residuals from the cover's low-intensity regions, achieving a significant PSNR gain of +4.66 dB for secret images at 10 dB SNR while maintaining optimal fidelity when the noise is negligible. This pragmatic framework provides a robust solution for deploying INN-based steganography in realistic, lossy transmission scenarios.

키워드