ReViS: Closed-loop invertible video steganography via redundant prior aggregation

Existing video steganography methods based on invertible neural networks (INNs) suffer from severe redundant prior information loss during forward hiding. Relying on unguided Gaussian sampling for backward extraction disrupts network bijectivity, causing spatiotemporal inconsistency, temporal drift, and vulnerability to steganalysis and active channel distortions (e.g., video compression). To overcome these challenges, we propose Reversible Video Steganography (ReViS), a closed-loop framework leveraging redundant prior aggregation. ReViS utilizes the discrete wavelet transform (DWT) for frequency-domain decoupling and integrates three synergistic modules. First, a wavelet-driven 3D channel attention (WD-3DCA) mechanism utilizes a hybrid architecture comprising a trainable light-weight fusion layer and a parameter-free 3D energy evaluation core to generate spatial masks from high-frequency subbands, guiding content-adaptive hiding without altering bijectivity. Second, a redundant prior aggregation module (RPAM) compresses discarded high-dimensional features into compact global priors, implicitly embedding them into robust low-frequency subbands. Finally, a frequency-aware redundancy prediction module (FA-RPM) provides a highly accurate, deterministic estimation of these variables at the receiver by fusing baseline features, priors, and spatial masks. Experimental results demonstrate that ReViS achieves high imperceptibility (stego PSNR of 45.52 dB) and near-lossless secret reconstruction (recovery PSNR of 48.41 dB). Furthermore, the framework exhibits robust statistical security against advanced spatiotemporal steganalyzers and maintains high reconstruction fidelity under severe physical distortions. Crucially, the DWT-based architecture significantly reduces memory and computational overhead, enabling a high throughput of 70 FPS to support real-time covert communication.