Time efficient joint optimization federated learning over wireless communication networks

Abstract

Artificial intelligence (AI) has made a profound impact on our daily life. The 6th generation mobile networks (6G) should be designed to enable AI services. The native intelligence is introduced as an important feature in 6G. 6G native AI network is realized by the philosophy of federated learning (FL) to ensure data security and privacy. Federated learning over wireless communication networks is treated as a potential solution to realize native AI. However, introducing FL in the 6G will lead to expansive communication cost and unstable FL convergence with unreliable air interface. In this paper, we propose a solution for FL over wireless networks and analyze the training efficiency. To make full use of the advantages of the proposed network, we introduce a communication-FL joint optimization (CFJO) algorithm by jointly considering the effects of uplink resource, energy consumption and latency constraints. CFJO derives a transmission strategy with resource allocation and retransmissions to reduce the wireless transmission interruption probability and model upload latency. The simulation results show that CFJO significantly improves the model training efficiency and convergence performance with lower interruption probability under the latency constraint.