Deep-learning-based ciphertext-only attack on optical scanning cryptosystem

Abstract

Optical scanning cryptography (OSC) has garnered considerable attention because of its ability to acquire an incoherent hologram from a physical object using a single-pixel sensor. Security analysis is crucial to improve the performance of cryptographic systems. In this study, we prove that the optical scanning cryptosystem is vulnerable to a deep-learning-based ciphertext-only attack (COA) strategy. The conventional algorithm denotes the decryption process as an inverse problem of the encryption process and directly trains the deep learning architecture with a large number of pairs of plaintext-ciphertext images. However, this study first presents a combined deep-learning-based framework dedicated to processing COA on OSC using three deep neural networks (DNNs). With these DNNs, the plaintext image can be retrieved from an unknown ciphertext image in real time. The proposed method is independent of plaintext-ciphertext image pairs acquired from the OSC system, and thus makes COA available. Further, the proposed deep-learning-based COA method can attack an optical scanning cryptosystem encrypted with either pure phase keys or complex amplitude keys. Simulations and experimental results proved the feasibility and effectiveness of the proposed deep-learning-based COA method.