Grid-Assisted Federated Learning in Vehicular Crowdsensing: A Mobility-Aware and Probabilistic Vehicle Selection Approach

Abstract

Recently, Federated Learning (FL) has been advanced for Vehicular Crowdsensing (VCS), termed F-VCS, to enable collaborative learning without sharing private data, thus promoting privacy. However, the mobility of Intelligent Connected Vehicles (ICVs) leads to dynamic changes among FL participants, becoming one of the main challenges limiting F-VCS performance, as these dynamics can affect model training stability and accuracy. Current research primarily addresses these dynamics through ICV selection and resource optimization mechanisms. However, due to the lack of an effective method for characterizing ICV mobility, these approaches often rely on making binary judgments based on certain constraints (e.g., learning delay) rather than comparing the ICVs' mobility characteristics. As a result, they may not yield the most efficient outcomes for adapting to vehicular network dynamics, affecting overall system performance. This paper introduces a <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</u>rid-<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A</u>ssisted <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">FL</u> framework for VCS, named VCS_GAFL, which measures ICV mobility using a grid-cell model. Unlike traditional binary selection methods, our approach utilizes participation probabilities to capture the mobility differences among individual ICVs, guiding more precise and effective ICV selection and enhancing the VCS system's overall performance and adaptability. Specifically, we introduce an MMP-V model to predict ICV participation probabilities in F-VCS tasks. We then derive expressions for the impact of ICV participation probabilities on expected FL loss and ineffective energy consumption and formulate an optimization problem to jointly optimize ICV participation probabilities, wireless resource allocation, and ICV selection to minimize these impacts. By solving this problem, we derive a closed-form solution for wireless resource allocation and propose a probability-guided ICV selection strategy. Experimental results demonstrate that VCS_GAFL significantly enhances global accuracy, highlighting its potential and value in improving F-VCS services.