MIRA: Multi-scale invertible dual-attention redundancy-aware network for high-capacity video steganography

Achieving high embedding capacity while maintaining visual imperceptibility and strict reversibility remains a critical challenge in deep video steganography. Most existing invertible neural networks suffer from limited feature representation and inadequate modeling of spatio-temporal redundancy. To address these limitations, we propose MIRA, a Multi-scale Invertible Dual-Attention Redundancy-Aware Network. The framework incorporates a Multi-scale Hybrid Invertible Block to enhance global-local feature fusion and introduces a Dual-Attention Redundancy Refinement Module (DARRM). Within DARRM, we synergize local convolutional attention with non-local sparse attention to explicitly capture spatio-temporal redundancies, enabling fine-grained, content-adaptive feature optimization. Extensive experiments on the Vimeo-90K, UCF-101, and DAVIS datasets demonstrate that MIRA consistently achieves state-of-the-art performance. On Vimeo-90K, our method yields a stego video PSNR of 45.25 dB, a recovered secret video PSNR of 48.45 dB, and a Temporal Consistency Error (TCE) of 0.98×10^-3, with a recovery Mean Absolute Error (MAE) of 0.68. Moreover, the proposed framework exhibits superior zero-shot generalization and comprehensive resilience against diverse real-world distortions and steganalysis attacks. This work provides a highly efficient, secure, and robust solution for practical video steganography.