We introduce the powerful Deep Learning (DL) strategy to evaluate the security strength of computational-ghost-imaging-based cryptosystems. It's well known that the Computational Ghost Imaging (CGI) technique has an excellent advantage of anti-noise ability, hence the CGI-based cryptosystems are much easier to be optically implemented, compared with previous various kinds of optical cryptosystems. However, an objective and comprehensive security analysis still must be carried out before any a cryptosystem is claimed to be safe or reliable. Unfortunately, we find out that the current CGI-based cryptosystems and even its security-enhanced ones both have security flaws. If an opponent gets a set of chosen plaintexts and their corresponding ciphertexts, an "equivalent key" or "equivalent decryption network" could be find out by training the pairs of "plaintext-ciphertext" with DL method, in which we select the Deep Neural Network (DNN) as the fundamental structure. Furthermore, we also carried out some simulations to demonstrate the performance of our proposed method under two severe conditions: down-sampling and Gaussian noise. We believe our methodology is suitable for a major part of existing optical cryptosystems including the CGI-based ones.
展开▼
