Integrating idle embedded devices into cloud computing is a promising approach to support Distributed Machine Learning (DML). In this paper, we approach to address the data hiding problem in such DML systems. For the purpose of the data encryption in DML systems, we introduce the tripartite asymmetric encryption theorem to provide theoretical support. Based on the theorem, we design a general image encryption scheme (called ArchNet), which can encrypt original images via the encoder to resist against illegal users. ArchNet encrypts the data set by a specific neural network, which is especially trained for encryption. The encrypted data can be easily recognized by deep learning model. However, the encrypted data can not be recognized by human, which makes the illegal attacker difficult to steal the encrypted data. We use MNIST, Fashion-MNIST and Cifar-10 datasets to evaluate efficiency of our design. We deploy certain base models on the encrypted datasets and compare them with the RC4 algorithm and differential privacy policy. Our design can improve the accuracy on MNIST up to 97.26% compared with RC4. The accuracies on these three datasets encrypted by ArchNet are similar to the base model. ArchNet can be deployed on DML systems with embedded devices.

将空闲的嵌入式设备集成到云计算中是一种支持分布式机器学习（DML）的有前途的方法。在本文中，我们将解决此类DML系统中的数据隐藏问题。为了DML系统中的数据加密，我们引入了三方非对称加密定理，以提供理论上的支持。基于该定理，我们设计了一种通用的图像加密方案（称为ArchNet），该方案可以通过编码器对原始图像进行加密，以抵制非法用户。ArchNet通过特定的神经网络对数据集进行加密，该网络经过了专门的加密培训。深度学习模型可以轻松识别加密的数据。但是，加密数据无法被人类识别，这使得非法攻击者难以窃取加密数据。我们使用MNIST，Fashion-MNIST和Cifar-10数据集可评估我们的设计效率。我们在加密的数据集上部署某些基本模型，并将其与RC4算法和差分隐私策略进行比较。与RC4相比，我们的设计可以将MNIST的准确性提高多达97.26％。ArchNet加密的这三个数据集的准确性与基本模型相似。可以将ArchNet部署在具有嵌入式设备的DML系统上。