Decentralized semi-federated learning in digital twin-empowered industrial IoT

Abstract

Recently, Industrial Edge Intelligence (IEI) is becoming a key component to achieve smart factory. However, the large amount of required training data, data privacy issues, and heterogeneous computation capacity among industrial devices all pose significant difficulties and challenges to the realization of IEI. In order to improve the efficiency of resource utilization and overcome the limitations of conventional federated learning-based IEI, we propose a Decentralized Semi-Federated Learning (DSFL) framework where data with higher privacy attributes remains on the local end and data with less privacy attributes can optionally upload to multiple potential edge nodes for training and constructing Digital Twins (DT). In order to minimize the training overhead of each round of the proposed DSFL framework, we investigate a novel perspective that the number of active edge nodes can be dynamically adjusted. Therefore, a new optimization problem for jointly optimizing edge association and non-private data uploading ratio is formulated. Considering that the formulated problem is a mixed integer programming problem, we propose efficient algorithms to find the solution. In addition, simulation results demonstrate that our proposed algorithm can significantly reduce the training overhead.