Empirical Analysis of Asynchronous Federated Learning on Heterogeneous Devices: Efficiency, Fairness, and Privacy Trade-offs

TL;DR

This work tackles the challenge of device heterogeneity in Federated Learning by comparing synchronous FedAvg and asynchronous FedAsync on a physical five-device edge testbed, with Local Differential Privacy enforced and privacy tracked via the Moments Accountant, using Speech Emotion Recognition as a privacy-sensitive benchmark. The study demonstrates that FedAsync can deliver up to $10\times$ faster convergence but incurs pronounced fairness and privacy disparities, with high-end devices dominating updates and accumulating greater privacy loss than low-end devices. These findings highlight the need for adaptive aggregation and privacy mechanisms that jointly balance efficiency, fairness, and privacy in realistic heterogeneous deployments. The results offer practical guidance for designing scalable, equitable, and privacy-preserving FL systems in real-world edge environments, and suggest promising directions toward joint aggregation–privacy adaptation and fairness-aware personalization.

Abstract

Device heterogeneity poses major challenges in Federated Learning (FL), where resource-constrained clients slow down synchronous schemes that wait for all updates before aggregation. Asynchronous FL addresses this by incorporating updates as they arrive, substantially improving efficiency. While its efficiency gains are well recognized, its privacy costs remain largely unexplored, particularly for high-end devices that contribute updates more frequently, increasing their cumulative privacy exposure. This paper presents the first comprehensive analysis of the efficiency-fairness-privacy trade-off in synchronous vs. asynchronous FL under realistic device heterogeneity. We empirically compare FedAvg and staleness-aware FedAsync using a physical testbed of five edge devices spanning diverse hardware tiers, integrating Local Differential Privacy (LDP) and the Moments Accountant to quantify per-client privacy loss. Using Speech Emotion Recognition (SER) as a privacy-critical benchmark, we show that FedAsync achieves up to 10x faster convergence but exacerbates fairness and privacy disparities: high-end devices contribute 6-10x more updates and incur up to 5x higher privacy loss, while low-end devices suffer amplified accuracy degradation due to infrequent, stale, and noise-perturbed updates. These findings motivate the need for adaptive FL protocols that jointly optimize aggregation and privacy mechanisms based on client capacity and participation dynamics, moving beyond static, one-size-fits-all solutions.

Figures (5)

• Figure 1: Federated learning under device heterogeneity. Comparison of synchronous and asynchronous FL for speech emotion recognition. Local update privacy is ensured by DP-SGD, with per-client privacy loss tracked by the moments accountant. Devices range from low-end (HW $T_1$) to high-end (HW $T_5$), reflecting realistic hardware variability.
• Figure 2: FL testbed with clients comprising heterogeneous devices (HW $T_1$--$T_5$)
• Figure 3: FL training performance per round across heterogeneous devices.
• Figure 4: Convergence time of FedAvg vs. FedAsync with and without staleness-aware aggregation under device heterogeneity.
• Figure 5: Global and per-device accuracy trajectories, along with participation percentages (PP), in asynchronous FL with staleness-aware weighting under varying aggregation strengths $\alpha$. Higher $\alpha$ accelerates convergence but exacerbates representational skew by diminishing participation and degrading accuracy on high-staleness clients.