Deep Hologram Prior Network for Polarization-Assisted Reconfigurable Metasurface Visual Secret Sharing

Polarization-assisted visual secret sharing (VSS), based on metasurfaces, provides a promising route toward secure and computation-free optical encryption by distributing secret information into orthogonal polarization channels. However, existing implementations are predominantly limited to passive metasurfaces, which fundamentally constrain dynamic reconfiguration and multisecret switching. Here, we present a reconfigurable dual-polarization metasurface platform, enabled by a deep hologram prior neural network, for physically constrained phase hologram optimization. By adopting a simplified 2 × 2 supercell architecture, the proposed tunable meta-unit achieves independent and stable phase control for both polarizations, with reduced system complexity. The deep hologram prior network further enables efficient hologram synthesis under different control complexities and improves the imaging quality by about 35% compared to the weighted Gerchberg-Saxton method. We experimentally demonstrate dynamic-polarization-assisted VSS using a reconfigurable metasurface prototype operating at 10 GHz, with results in good agreement with theoretical predictions. This work establishes a scalable and flexible framework for dynamic optical encryption and may find applications in secure communication, information authentication, and reconfigurable photonic security systems.