Abstract The so-called Optical Asymmetric Cryptosystems (OACs) have attracted more and more attention due to the unique mechanism of encryption/decryption and higher security level resulted from the involved nonlinear operations. Meanwhile, several attack methods have also been reported to analyze the security risk of typical OACs. In this work, we first stated our basic views about two famous OACs and then demonstrated that they are both vulnerable to a deep-learning-based strategy. Thanks to a carefully designed and trained (by known plaintext-ciphertext pairs) generative adversarial network (GAN), an attacker could intercept enough high-frequency components of subsequent plaintext leading to successful retrieval. Compared with the previous amplitude-phase-retrieval-based methods, the proposed learning-based scheme has a major advantage of retrieving plaintexts with high quality in real-time. Numerical simulations demonstrate the feasibility and effectiveness of the proposed learning-based attack method.

中文翻译：

---

一种基于学习的光非对称密码系统攻击方法

摘要 所谓的光学非对称密码系统(OACs)由于其独特的加解密机制和涉及非线性运算的更高安全级别而越来越受到人们的关注。同时，也报道了几种攻击方法来分析典型OAC的安全风险。在这项工作中，我们首先陈述了我们对两个著名 OAC 的基本观点，然后证明它们都容易受到基于深度学习的策略的影响。由于精心设计和训练（通过已知的明文-密文对）生成对抗网络 (GAN)，攻击者可以拦截后续明文的足够高频分量，从而成功检索。与以往基于幅相检索的方法相比，所提出的基于学习的方案具有实时检索高质量明文的主要优势。数值模拟证明了所提出的基于学习的攻击方法的可行性和有效性。