Deep learning iterative optical asymmetric cryptosystem model based on QR decomposition and diffractive imaging

Deep learning techniques not only facilitate the analysis of vulnerabilities in optical cryptosystems but also serve as a foundation for optical cryptography. In the proposed cryptosystem, multiple grayscale images are encrypted using orthogonal-triangular (QR) decomposition within the Fresnel transform domain. The resulting ciphertexts are captured at three distinct axial positions using a digital sensor. The ciphertexts corresponding to different input images, are then stored within a deep learning model. For decryption, a novel pre-trained deep learning model other than conditional generative adversarial network has been developed which is based on an iterative algorithm. During training, the neural network learns to establish a mapping between the three different ciphertexts and their respective original images. The decryption process requires both, the pre-trained model parameters and asymmetric keys derived from QR decomposition, ensuring that security is not solely dependent on the neural network. To the best of our knowledge, this is the first implementation of a deep learning model integrated with QR-based diffractive imaging. By combining QR-decomposed keys, varying Fresnel distances, and deep learning-based key integration, the proposed system enhances security and demonstrates resilience against known-plain text attack.