FedAsyncM: A Hybrid Asynchronous Federated Learning Aggregation Algorithm

Federated Learning (FL) enables distributed model training across edge devices without sharing raw data. However, client heterogeneity and asynchronous updates introduce challenges in convergence and accuracy. In this paper, we use the APPFL framework (Advanced Privacy-Preserving Federated Learning) to compare eight FL aggregation algorithms— including synchronous methods (FedAvg, FedAvgM, FedAdam, FedYogi, FedAdagrad) and asynchronous methods (FedAsync, FedBuffer, and our proposed FedAsyncM)—on an MNIST classification task using a convolutional neural network. We propose a novel algorithm, FedAsyncM, which adds a server-side momentum term to the FedAsync aggregator. Our experiments, conducted on a homogeneous network of Raspberry Pi 5 devices and a heterogeneous mixed cluster including a faster laptop, were later extended to simulations of up to 100 clients to measure test accuracy and training time. The results show that the synchronous FedAvgM achieves the highest accuracy, while the asynchronous FedAsync achieves the fastest training time. FedAsyncM, with carefully chosen parameters (e.g., learning rate and momentum), balances this trade-off by converging nearly as quickly as FedAsync while attaining accuracy close to FedAvgM. These findings indicate that incorporating momentum into asynchronous FL can significantly improve performance.