The conventional diffractive-imaging-based image encryption (CDIE) proposed in 2010 has drawn much attention in the last decade. A new modified diffractive-imaging-based image encryption (MDIE) with good non-linearity was proposed in 2019. Inspired by the cryptoanalysis method based on the convolutional neural networks, we propose a model trained with large numbers of ciphertext-plaintext pairs to crack the new cryptosystem. The proposed model has two blocks, the first block is employed for extracting the features of ciphertext and the second one is used for recovering the plaintext according to these extracted features. Compared with existing cryptoanalysis methods based on the convolutional neural networks, the proposed model has better generalization. We hope this structure can help researchers to solve other optical cryptoanalysis problems. To our knowledge, this is the first time to make the CNN-based attack can be used beyond the MNIST dataset, including both the handwriting dataset and fashion dataset. This work proves the CNN-based attack can be used in general situations. The analysis of validity and robustness is presented. Besides, the experimental results are conducted to validate the proposed method.

中文翻译：

---

基于深度学习攻击的改进的基于衍射成像的图像加密的密码分析

2010 年提出的传统基于衍射成像的图像加密（CDIE）在过去十年中引起了广泛关注。2019 年提出了一种新的改进的基于衍射成像的图像加密（MDIE），具有良好的非线性。受基于卷积神经网络的密码分析方法的启发，我们提出了一个用大量密文-明文对训练的模型来破解新的密码系统。所提出的模型有两个块，第一个块用于提取密文的特征，第二个块用于根据提取的特征恢复明文。与现有的基于卷积神经网络的密码分析方法相比，该模型具有更好的泛化性。我们希望这种结构可以帮助研究人员解决其他光学密码分析问题。据我们所知，这是第一次使基于 CNN 的攻击可以在 MNIST 数据集之外使用，包括手写数据集和时尚数据集。这项工作证明了基于 CNN 的攻击可以用于一般情况。提出了有效性和稳健性的分析。此外，还进行了实验结果以验证所提出的方法。