A research team has demonstrated an asynchronous federated learning framework called FedQS that mitigates the combined challenges of outdated model updates and participation imbalances across heterogeneous computing environments. In distributed machine learning networks spanning the Internet of Things and edge devices, training must proceed without pooling private raw data on a central server. While traditional synchronous federated learning forces the central server to wait for the slowest participating device before performing model aggregation, asynchronous paradigms allow clients to transmit local model updates independently. However, asynchronous operation introduces severe complications, notably model staleness when updates arrive late based on older global model states, as well as training bias when fast devices dominate the aggregated model while slower devices are marginalized. Across benchmark evaluations, the newly proposed system achieved an average accuracy improvement of 8.62 percent over four established asynchronous baselines.
The study, published in the Journal of Cloud Computing by researchers Jiahui Zhou, Fang Li, Tianyu Shi, and Xingxiao Quan, identifies a critical structural limitation in previous asynchronous distributed optimization. When client devices operate across diverse hardware profiles and unstable wireless connections, local data distributions become highly non-uniform and non-identically distributed. This statistical heterogeneity causes client drift, pulling local optimization steps in conflicting directions and destabilizing global inference accuracy. Existing asynchronous methods frequently discard outdated parameters or prioritize fast updates to maximize system throughput, yet discarding slow clients suppresses their unique data distributions and exacerbates training bias. Conversely, incorporating severely outdated gradient calculations injects gradient deviations that hinder mathematical convergence. FedQS addresses this trade-off by decoupling the management of model staleness from client participation through three coordinated mechanisms: dynamic staleness evaluation, helper-based queue scheduling, and contribution-calibrated aggregation.
To evaluate parameter updates with greater precision than simple communication latency measurements, FedQS implements a multi-dimensional staleness assessment. The central server tracks the end-to-end duration of each client by computing the difference between the timestamp when a model was dispatched and when the updated weights were returned. Over historical update rounds, the server maintains an average delay metric for each client and derives a dynamic median threshold to divide participating devices into fast and slow categories. When an update arrives, its freshness is determined not solely by wall-clock time, but by combining a cosine-distance divergence calculation between local and global parameters with a delay-weighted logarithmic scaling factor. Because larger parameter divergence indicates greater staleness, the cosine-distance metric penalizes updates that diverge substantially from the current global parameter trajectory. Updates exhibiting a calculated score exceeding the dynamic average staleness of the active comparison window are marked as high-staleness records, while the rest are classified as low-staleness records.
Following classification, uploaded model updates are systematically organized into four distinct server queues: low-staleness updates from fast clients, high-staleness updates from fast clients, low-staleness updates from slow clients, and high-staleness updates from slow clients. Instead of holding bare numerical parameter vectors, each queue entry is structured as a rich metadata record storing parameter weights, a constituent list of atomic records, the calculated staleness score, processing status, source delay or version, and a monotonically increasing enqueue index. High-staleness records submitted by fast clients can be combined internally once a designated queue threshold is reached, after which their combined staleness is re-evaluated against the latest global reference model before they can transition into an aggregation-ready candidate queue. Meanwhile, low-staleness updates from both fast and slow devices are routed directly toward global aggregation eligibility.
The central innovation of the framework lies in its queue-based scheduling algorithm, which resolves high-staleness updates from slow clients without compromising data confidentiality. When a slow device produces an update that suffers from high staleness, the server enlists an available fast client to act as a computational helper. To qualify as an active helper, a fast device must have recently uploaded a low-staleness update, must not be currently engaged in another reprocessing assignment, and must be selected based on the earliest enqueue timestamp. Rather than transferring any private raw data, samples, labels, or feature loaders from the original client, the central server dispatches only the stale model parameters alongside task metadata to the designated helper. The helper client then initializes its local optimization routine from those stale parameters and conducts continuation training exclusively on its own local, private training data before returning the reprocessed parameters to the server.
Crucially, the server maintains rigorous provenance tracking throughout this helper reprocessing workflow to safeguard training fairness. When the helper completes additional local training passes and transmits the reprocessed parameter vector back to the server, the server preserves the original slow client’s identification tag in the constituent record rather than reassigning credit to the helper. This scheduling provenance ensures that the contribution of the slow device remains recognized during future aggregation stages, preventing frequent fast contributors from overwhelming the optimization path. The helper device’s completion does not increment its own cumulative submission counter. Before the reprocessed update can enter the global aggregation queue, the server re-evaluates its staleness score against the most current version of the global model; if the recalculation confirms that staleness has fallen below the running threshold, the update qualifies for inclusion in global aggregation.
Global aggregation is deterministic and governed by a predefined threshold representing a normalized low-staleness aggregation proportion. Whenever the cumulative number of validated records in the low-staleness queues satisfies this threshold, the server extracts the earliest eligible updates based on their global enqueue index. The aggregation weighting formula actively counteracts fast-client dominance by scaling each constituent’s weight proportionally to its original client’s local dataset size divided by the natural logarithm of its cumulative submission count plus one. By dampening the influence of clients that submit frequent updates, the adaptive formula protects the global model from being steered disproportionately toward the statistical distributions of fast hardware. The server then computes the weighted aggregate, executes a relaxation update step against the global model, and broadcasts the updated global parameters back to the associated original participating devices.
Theoretical convergence analysis of a relaxed record-level abstraction of FedQS demonstrates that the framework satisfies a standard nonconvex optimization descent bound. Under standard analytical assumptions including objective smoothness, bounded stochastic gradient variance, bounded model drift, and bounded effective staleness enforced by queue scheduling, the theoretical derivation confirms that the average squared gradient norm diminishes with the total number of communication iterations. The rate analysis shows that model aggregation errors remain explicitly bounded by helper-induced effective source mismatches, bookkeeping deviations, and residual abstraction factors, providing formal structural support for the algorithm’s empirical stability under non-uniform device and data conditions.
Empirical validation of the framework was conducted within the easyFL and FLGo federated learning experimental platforms across 100 simulated artificial intelligence of things devices using convolutional neural network and ResNet-18 architectures. Training was executed across varying proportions of slow devices and latency disparities ranging from tenfold to one-hundredfold differences on the Fashion-MNIST and CIFAR-10 image classification benchmarks under non-identical Dirichlet data partitioning. Across all tested latency and slow-client ratios, FedQS consistently surpassed the four evaluated baseline frameworks: FedAsync, FedBuff, KAFL, and CA²FL. On Fashion-MNIST, FedQS achieved 81.26 percent test accuracy under a thirty percent slow-client ratio with a tenfold latency gap, outperforming the closest baseline, CA²FL, at 78.13 percent. When the latency penalty increased to one hundred times, FedQS exhibited minimal degradation, dropping slightly to 80.92 percent, whereas FedAsync declined from 71.11 percent to 68.58 percent.
Similar performance resilience was recorded on the more complex CIFAR-10 benchmark using the ResNet-18 network. In tests where slow clients comprised thirty percent of the fleet with a tenfold latency difference, FedQS attained 58.28 percent accuracy compared to 54.59 percent for CA²FL. Under severe conditions where slow clients represented seventy percent of the network and operated with a one-hundredfold latency lag, FedQS maintained 53.35 percent accuracy, whereas FedBuff and KAFL fell to 44.56 percent and 42.21 percent, respectively. Throughout these multi-seed benchmark evaluations, FedQS maintained standard deviations between ±0.11 and ±1.28 percent, illustrating substantially lower training volatility compared to baseline methods whose standard deviations fluctuated as high as ±9.36 percent.
Component-wise ablation experiments on Fashion-MNIST, CIFAR-10, and CIFAR-100 isolated the practical contributions of staleness evaluation, queue scheduling, and contribution-based aggregation. Removing queue scheduling caused the most severe performance penalty across the three datasets, reducing average accuracy across the final evaluation window by 20.14 percentage points and confirming that helper-assisted parameter reprocessing is essential for suppressing stale update disruption. Removing contribution-based aggregation decreased cross-dataset accuracy by 8.85 percentage points, while eliminating cosine-distance staleness evaluation produced a 4.33 percentage point reduction. Although the staleness evaluation component exhibited slight dataset sensitivity, the three mechanisms together demonstrated strong collective complementarity, delivering an overarching architectural solution for robust asynchronous edge learning.
Subject of Research: Asynchronous federated learning and queue scheduling in heterogeneous distributed computing
Article Title: FedQS: asynchronous federated learning based on queue scheduling
Article References: Zhou, J., Li, F., Shi, T., & Quan, X. (2026). FedQS: asynchronous federated learning based on queue scheduling. Journal of Cloud Computing, 15(1), Article 125. https://doi.org/10.1186/s13677-026-00977-4
Image Credits: AI Generated
DOI: 10.1186/s13677-026-00977-4
Keywords: asynchronous federated learning, FedQS, queue scheduling, model staleness, training bias, edge computing, Internet of Things, client drift, non-IID data, distributed machine learning, Fashion-MNIST, CIFAR-10, CIFAR-100, ResNet-18, convolutional neural networks, helper clients, cosine distance
News Source: Veronica Carney. (November 19, 2026). Asynchronous Federated Learning Framework Uses Queue Scheduling to Improve Model Training Accuracy. Scienmag.



