The Impact Analysis of Delays in Asynchronous Federated Learning With Data Heterogeneity for Edge Intelligence

Federated learning (FL) has provided a new methodology for coordinating a group of clients to train a machine learning model collaboratively, bringing an efficient paradigm in edge intelligence. Despite its promise, FL faces critical challenges in Internet of Things (IoT) networks, particularly the combined impact of data heterogeneity and communication delays. This paper examines the theoretical and empirical impact of these factors in Asynchronous Federated Learning (AFL). Initially, we categorize existing memoryless asynchronous strategies as Asynchronous Updates with Delayed Gradients (AUDG). Our theoretical analysis of AUDG reveals the coupling effect of delays exacerbate the adverse impact of data heterogeneity, causing the global model to drift towards frequently active clients. To address this, we propose a gradient reusing mechanism, termed Pseudo-Synchronous Updates by Reusing Delayed Gradients (PSURDG). By leveraging storage to reuse historical gradients, PSURDG effectively decouples the correlation between delay and data heterogeneity. Crucially, we conducted a comprehensive convergence analysis covering both convex and non-convex settings, confirming the algorithm’s effectiveness in diverse optimization landscapes. Finally, both schemes are validated through rigorous analysis and extensive simulations on multiple datasets. The results demonstrate a clear trade-off that AUDG remains efficient under low data heterogeneity, while PSURDG improves convergence in high-heterogeneity scenarios with moderate delays, thereby providing a theoretical guideline for aggregation strategy selection in different edge intelligence scenarios.