Edge aggregation placement for semi-decentralized federated learning in Industrial Internet of Things

Abstract

Rapid technological advancements have resulted in more smart devices with an increased volume of data in Industrial Internet of Things (IIoT) systems. These data often contain significant levels of heterogeneity and privacy in accordance with the various requirements of industrial applications, therefore controlling for these data in an intelligent and secure way is still an outstanding problem. Federated edge learning (FEEL), a distributed machine learning paradigm, has become a viable option for developing learning models while maintaining data privacy. In this paper, we concentrate on a semi-decentralized FEEL (SD-FEEL) framework, taking into account the limited training data in a single central cluster, where the shared edge models is trained among devices and edge servers. An edge aggregation optimization problem with joint device association, resource block allocation, and edge server placement is then formulated based on the evaluated upper bound of the convergence performance for SD-FEEL, with the goal of minimizing the training loss while taking into account the cost budget of edge servers. Based on the calculation of training loss degradation, the proposed optimization issue can be transformed into a dynamic optimum problem. A Trilateral Matching-based association (TMA) method is then proposed to discover solutions to the sub-problems of device association and resource block allocation. Additionally, a Tabu Search-based Service location (TSP) method is suggested to optimize the placement of the edge servers. To fully address the optimization problem, a combination iterative approach including the TMA and TSP is proposed. According to simulation data, the proposed algorithm can greatly beat the baselines in terms of test accuracy with limited cost budget.