A Fair and Rational Data Sharing Strategy Toward Two-Stage Industrial Internet of Things

Abstract

The easy and pervasive involvement of devices in Industrial Internet of Things has greatly benefited the implementation and adoption of various smart services. One prominent prerequisite of such trends is the extensive and continuous support and sharing of data and resources among devices. However, previous efforts usually treat the data sharing as one-time task among devices, which are incapable when the data are applied for the distributed and iterative training task of machine learning models. Therefore, this article proposes a novel framework for continuous data sharing in Industrial Internet of Things. The system consists of different system owners, each brings devices and participate the distributed training of models. Specifically, system owners hold different scales of devices, data, and resources, while devices own heterogeneous availability in different time periods. In this case, the goal is to properly assign devices for qualified model training process in different rounds, such that no devices will devote unlimited resources and the overall efforts and consumptions among different owners are balanced. Accordingly, three algorithms for device allocation are proposed, based on whether the availability of devices in each training round are known at the beginning of the training procedure. The analysis shows that all algorithms can achieve a rational allocation for devices and balance the performance among system owners. Finally, evaluation results reveal that the proposed solutions outperform baseline methods in providing better data sharing plans.