Secure multiparty computation is a popular technique in privacy preservation of machine learning. It allows a set of parties to jointly compute a function over their inputs while keeping them private. However, recent research on MPC-based privacy-preserving machine learning mainly focuses on the two-party setting and is considered slow when deployed in the multiparty setting. The privacy-preserving deep neural network training (PpNNT) we introduce in this article overcame these limitations. With relatively small security performance improvements compared with single-party setting and general multiparty computation approaches, the PpNNT is ready to support privacy-preserving computation in a wide variety of sensitive IIoT applications, including automatic assembly line, intelligent monitoring, digital transformation, etc. It could offer a practical and secure bridge of deep neural network training for some sensitive data islands.

By leveraging smart devices [e.g., industrial Internet of Things (IIoT)] and real-time data analytics, organizations, such as production plants can benefit from increased productivity, reduced costs, enhanced self-monitoring, and autonomous decision-making. In such a setting, machine learning plays an important role in data analytics, but the use of conventional centralized machine learning solutions may raise uncomfortable concerns about data privacy. Hence, one can explore the use of federated learning. In this article, we propose privacy-preserving deep neural network training (PpNNT), which is designed to support federated learning in the multiparty setting. To minimize the overall costs, we further design a hybrid architecture to fully maximize resource utilization. Our proposed design allows the PpNNT system to provide high security, efficiency, and scalability for IIoT data analytics, as evidenced by our theoretical security proof and experimental results on the CIFAR10 dataset.