Optical visual encryption using focused beams and convolutional neural networks

The target of this paper is to implement an optically-based visual encryption system able to work with a large set of optical codes. The optical setup comprises a holographic system designed to generate spirally-polarized highly focused fields and an imaging module able to perform polarimetric analysis. In a previous stage, the optical system is numerically simulated in order to produce synthetic polarimetric distributions that are used to train a convolutional neural network. Interestingly, the way the network is trained depends on the selected state of polarization. Then, secret codes are split in two XOR-connected ones that are optically processed. The corresponding experimental polarimetric distribution is obtained and transmitted to the corresponding recipients, that can recover the code by interrogating the neural network. Finally, combining the two pieces of information, the encrypted message can be decoded.