Optical image encryption based on two-channel detection and deep learning

Abstract

Aside from its application to the vulnerability analysis of optical cryptosystems, deep learning (DL) technique can also be used as a potential basis for optical cryptography itself. Here, we report a new optical image encryption scheme based on two-channel detection and DL. The encryption process consists of two-channel signal detection within the optical encryption system where three random binary masks are involved, and the training of a neural network to derive an approximate model. The two optically encoded patterns collected with a pair of CCD cameras are recombined into a single speckle image by using two independent sparse matrices. In the training section, the neural network is trained to find the mapping relationship between the speckle images and their corresponding original objects. Besides the model parameters, the decryption requires the use of the matrix keys, which avoids a situation where the DL-based cryptosystem is designed such that security relies entirely on the trained network itself. To the best of our knowledge, this is the first report using a single-wavelength, dual-channel system for diffuse optical encryption. Experimental results demonstrate the feasibility and good performance of the proposed scheme.