All-optical color image encryption using multimode fiber speckles and diffractive deep neural networks

To enable high-speed and secure color image transmission without the computational burden of deep learning, we present an all-optical encryption system that integrates spatial degree-of-freedom encoding of orbital angular momentum (OAM) beams, multimode fiber (MMF) speckle encryption, and diffractive deep neural network (D2NN) decoding. In the proposed scheme, the grayscale values and color channels of image pixels are respectively mapped to the topological charges and spatial positions of OAM beams, which are transmitted through a 1-meter MMF under open environmental conditions. During transmission, random disturbances such as vibrations and temperature-humidity fluctuations introduce speckle-based ciphertexts. At the receiver, two decoupled D2NN modules collaborate in parallel to decode the encrypted speckles. After training on speckle samples collected under long-term noise interference, the pre-trained D2NNs achieve an accuracy of 98.75% in RGB channel decryption and 96.53% in grayscale value recovery. For 512 × 512 color images, the end-to-end decrypted results exhibit structural similarity index measure (SSIM) values above 0.95. These results demonstrate that the proposed OAM-MMF-D2NN encryption framework achieves ultra-high fidelity and strong robustness, laying the foundation for computation-free, interference-resilient, and ultra-fast all-optical image encryption systems.