ArchNet: A data hiding design for distributed machine learning systems

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