Speckle-based cryptosystem with physically associated authentication and task-dedicated model matching

Existing deep-learning-enabled optical scattering encryption schemes lack effective task-level access control, preventing independent multi-task encryption and decryption. Furthermore, these schemes typically treat the trained single neural network as a global decryption key, introducing critical security vulnerabilities once the model is leaked. To achieve secure multi-task optical scattering encryption and reliable task-model matching, we propose a dual-channel scheme integrating physical marker authentication and a residual-network matching module. Task-specific physical signatures are created by the spatial positions of scattering media and dedicated marker images, while complementary binary matrices serve as digital keys, establishing a hybrid physical-digital access control system. After that, the optical setup generates a marker speckle and an interference-fused speckle carrying both plaintext and marker information; the final ciphertext is constructed by extracting pixels from the corresponding speckle pairs using customized digital keys. Decryption follows two-level verification: physical authentication using the correct binary matrix, followed by activation of the corresponding task-specific residual network for high-quality reconstruction. Experiments demonstrate reliable task-model matching across multiple independent task groups, robust security against key leakage and unauthorized decryption, and effective resistance to cross-verification attacks. This work establishes, what is believed to be, a new paradigm with strengthened physical security for multi-task optical scattering encryption.