Known-plaintext cryptanalysis for a computational-ghost-imaging cryptosystem via the Pix2Pix generative adversarial network

Abstract

A neural network based known-plaintext cryptanalysis for a computational-ghost-imaging (CGI) cryptosystem is proposed, which utilizes relevant physical priors as references and conditions during training. It retrieves more significant qualified and higher accurate attacking results with much more complicated human face dataset that fairly improves previous learning based works. Popularly employed neutral networks such as the convolutional neural network (CNN), recurrent neural network (RNN) and U-Net are further compared. However, our proposed method defeats them with the best attacking results, which is also proved by the following detailed quantitative analysis. On the other hand, compared with traditional methods utilizing phase recovering to estimate the privacy key, the proposed network method develops an end-to-end way that directly converts the ciphertext to the corresponding plaintext. The method is capable of high-volume attacking mission with rather highly qualified attacking results and fast response after valid training. Both computer simulations and optical experiments demonstrate the feasibility and effectiveness of the proposed method.