In response to increasingly stringent data privacy regulations, federated learning (FL) has gained prominence for its privacy-preserving characteristics. To implement FL intelligence efficiently, researchers have turned to mobile edge computing (MEC). However, existing efforts often overlook practical challenges in MEC systems, such as device heterogeneity, unstable channel conditions, and unpredictable user mobility, any of which, if mishandled, can lead to critical failures in FL. This paper introduces a novel FL framework, Multicore FL (MC-FL), designed to enable the successful deployment of FL intelligence in realistic MEC systems.
MC-FL stands out for its ability to cater to heterogeneous devices by maintaining and training multiple global models (GMs) with varying tradeoffs between learning performance and computational complexity. Notably, it introduces a partial client participation scheme to handle random client availability, enabling adaptability in uncertain mobile environments. The paper rigorously proves the convergence of MC-FL by bounding the performance gap between GMs learned by MC-FL and optimal GMs. These convergence results offer valuable insights into selecting the number of global and local training rounds, providing theoretical performance guarantees for MC-FL in realistic MEC systems. To enhance the training process of MC-FL, the paper proposes an online client scheduling scheme aimed at minimizing completion time. Leveraging a graph representation considering the priority of FL training tasks and time-varying wireless transmission rates, this dynamic scheduling scheme utilizes an event-triggered control policy, ensuring adaptability over the continuous timeline. In addition to theoretical advancements, the paper establishes a service provisioning scenario for MC-FL. In this scenario, service subscribers strategically download ready-to-use GMs from the edge server to process local service tasks. The service provisioning problem considers factors such as inference accuracy, service delay, device energy consumption, and configuration cost, aiming to maximize Quality of Experience (QoE) for subscribers. Experimental results demonstrate that MC-FL provides flexible service delivery, significantly improving subscribers’ QoE.
In summary, this paper introduces MC-FL as a comprehensive solution for FL in MEC systems, offering a robust framework, theoretical foundations, efficient training methodologies, and practical insights into real-world service provisioning scenarios. MC-FL emerges as a promising approach to address the challenges posed by privacy concerns, device heterogeneity, and dynamic mobile environments in the context of federated learning and edge computing.
Part of this work is supported by National Science Foundation.
Y Bai, L Chen, J Li, J Wu, P Zhou, Z Xu, J Xu, IEEE internet of things journal 10 (7), 5940-5952, 2022
https://ieeexplore.ieee.org/abstract/document/9961868
