Personalized Asynchronous Federated Learning for Intelligent Vehicular Computing

Smart consumer electronics, especially Intelligent Vehicular Computing (IVC), have surged in popularity. The influential technologies driving this popularity include the Internet of Things (IoT) and Artificial Intelligence (AI), empowering IVC with intelligent service. However, the popularization of IVC has brought a range of issues. IVC requires massive amounts of user data, often relying on federated learning (FL) framework to train AI models, facilitating for driving tasks such as road sign recognition. However, existing FL frameworks cannot adapt to vehicle processing delays due to the limitations of synchronous mechanisms or the lag effects on model update caused by asynchronous mechanisms. Furthermore, since vehicles are often in different environments, the heterogeneous data collected by their sensors also poses challenges to model performance. Therefore, this paper proposes a novel <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">P</u>ersonalized <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">A</u>synchronous <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Fed</u>erated learning (PA-Fed) framework for IVC to address the aforementioned issues. We employ vehicles to deploy personalized models locally to learn local data representation, thereby eliminating the negative impact of data heterogeneity on model performance. Additionally, we employ a staleness and model difference-aware dynamic aggregation method on the server to mitigate data heterogeneity and the lag effect caused by strugglers. Simulation results show that PA-Fed achieves better performance compared to synchronous and asynchronous baselines.