Distribution-Aware Mobility-Assisted Decentralized Federated Learning

TL;DR

This work investigates mobility in decentralized federated learning (DFL) and demonstrates that introducing mobile clients can significantly boost model accuracy by improving information flow in sparse networks. It introduces two distribution-aware mobility patterns, Distribution-Aware Mobility (DAM) and Distribution-Aware Cluster-Center Mobility (DCM), which guide mobile clients to locations that mitigate data heterogeneity and connectivity sparsity, under a mobility constraint $R_m$. Through extensive experiments on MNIST and CIFAR-10 across varying network sizes and heterogeneity levels, DAM and especially DCM consistently outperform baseline Static and Random Movement, with gains up to about 8% in highly heterogeneous settings. The results underscore the practical value of mobility in DFL and pave the way for theoretical convergence analyses and relaxation of distribution/location knowledge assumptions.

Abstract

Decentralized federated learning (DFL) has attracted significant attention due to its scalability and independence from a central server. In practice, some participating clients can be mobile, yet the impact of user mobility on DFL performance remains largely unexplored, despite its potential to facilitate communication and model convergence. In this work, we demonstrate that introducing a small fraction of mobile clients, even with random movement, can significantly improve the accuracy of DFL by facilitating information flow. To further enhance performance, we propose novel distribution-aware mobility patterns, where mobile clients strategically navigate the network, leveraging knowledge of data distributions and static client locations. The proposed moving strategies mitigate the impact of data heterogeneity and boost learning convergence. Extensive experiments validate the effectiveness of induced mobility in DFL and demonstrate the superiority of our proposed mobility patterns over random movement.

Figures (5)

• Figure 1: DFL schematic with a UAV as a mobile client and seven fixed clients. It illustrates the local data distributions of each client and the UAV’s movement across three locations (L1, L2, and L3). The shaded regions represent communication ranges, and the aggregated data distribution at each UAV location is shown at the bottom for reference.
• Figure 2: Accuracy of DFL system on MNIST Dataset after 1000 rounds with 20 clients.
• Figure 3: Impact of $|\mathcal{C}_m|$, $R_c$ and $R_m$ on accuracy of DFL system on MNIST Dataset after 1000 rounds with 20 clients.
• Figure 4: Accuracy of DFL with 50 clients after 1000 rounds in a larger network with 5 mobile clients using MNIST dataset.
• Figure 5: Accuracy of DFL with 20 clients after 600 rounds in a network with 5 mobile clients using CIFAR-10 dataset.