Communication and Computation Efficient Federated Learning for Internet of Vehicles With a Constrained Latency
Shengli Liu, Guanding Yu, Rui Yin, Jiantao Yuan, Fengzhong Qu
Abstract
Considering the privacy and security issues in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Internet of Vehicles</i> (IoV), wireless <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">federated learning</i> (FL) can be adopted to facilitate various emerging vehicular applications. However, wireless FL would suffer from a large learning latency due to the limitation of bandwidth and computing power as well as the unreliable communication caused by vehicle mobility. To cope with these challenges, a new structure is designed in this paper to facilitate the implementation of FL for IoV. First, we apply the gradient compression and mini-batch federated stochastic gradient descent to reduce the local gradient transmission and computation. Then, with theoretical analysis of the convergence rate and the learning latency, the learning performance can be improved by maximizing the convergence rate under a constrained latency. Accordingly, an optimization problem is formulated to jointly optimize compression ratio, batch size, and spectrum allocation. To solve this problem, an iterative algorithm is developed by problem decomposition. From the results, compression ratio and batch size should be adjusted according to the channel state information and computing power of the road side units to boost the learning efficiency at the cost of slight degradation on the learning accuracy. The superiority of the proposed algorithm is finally demonstrated through extensive simulations.