⚡ electrical engineering

Online-Score-Aided Federated Learning for Resource-Constrained Wireless Clients with Continual Data Arrival

This paper proposes Online-Score-Aided Federated Learning (OSAFL), a novel algorithm designed to address the challenges of resource-constrained wireless clients with limited storage and continual data arrival by theoretically analyzing convergence bounds under various constraints and optimizing global aggregation weights to minimize errors.

Original authors: Ferdous Pervej, Minseok Choi, Andreas F. Molisch

Published 2026-07-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ferdous Pervej, Minseok Choi, Andreas F. Molisch

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a world where your smartphone, your smartwatch, and your car are all trying to learn a new skill together, like recognizing different types of clouds or predicting traffic jams. They don't want to send all their private photos or location history to a giant central computer because that feels like handing over your diary to a stranger. Instead, they use a clever trick called Federated Learning. Think of it like a group project where everyone studies their own notes at home, writes down their best ideas, and only sends those ideas to a teacher who combines them into a master guide. The original notes stay safe at home.

But here's the catch: in the real world, these devices aren't perfect. They have tiny batteries, limited storage space, and they are connected by wireless signals that can be spotty and slow. Furthermore, the world doesn't stand still. New data is constantly pouring in—like a new cloud forming or a new traffic pattern emerging—while old data gets deleted to make room. Most computer learning methods assume the data is a static pile of rocks that never changes, which is great for a museum but terrible for a living, breathing wireless network. This paper tackles the messy reality of teaching machines when the data is constantly shifting and the devices are running on fumes.

The researchers behind this study, Ferdous Pervej, Minseok Choi, and Andreas F. Molisch, realized that the standard way of doing this group project was failing in these tricky conditions. They noticed that if you just ask every device to send its ideas at the same time, some devices might be too busy or too low on battery to participate, while others might be sending ideas based on data that is already outdated. To fix this, they invented a new method called Online-Score-Aided Federated Learning (OSAFL).

Imagine the central teacher (the server) is running a chaotic classroom where students (the devices) are constantly swapping out their textbooks for new ones. In the old system, the teacher would just average everyone's answers, regardless of whether a student was actually paying attention or if their textbook was from last year. OSAFL changes the game by giving the teacher a special "scorecard." Before the teacher combines the answers, they calculate a score for each student based on how much they actually trained, how reliable their connection is, and how similar their new ideas are to the group's current direction. If a student's data is too old or their connection is too shaky, they get a lower score, meaning their answer counts for less in the final mix. This ensures the "master guide" learns from the most relevant and reliable sources at that exact moment.

The paper doesn't just guess that this works; the authors built a complex mathematical model to prove that this scoring system helps the group learn faster and more accurately, even when the data is messy and the devices are struggling. They then tested this idea in a massive computer simulation. They set up a virtual world with 25 different devices, using three popular image-recognition datasets (CIFAR10, Fashion-MNIST, and MNIST) and three different types of AI models (a small one called SqueezeNet, a medium one called CNN, and a large one called ResNet-18).

In these simulations, the new OSAFL method consistently outperformed the existing "modified" versions of standard learning algorithms. For instance, when using the small SqueezeNet model on the CIFAR10 dataset, OSAFL achieved a test accuracy of about 41.7%, while the next best method only reached about 37.2%. On the Fashion-MNIST dataset, the gap was even wider, with OSAFL hitting 81.5% accuracy compared to roughly 78.8% for the others. The simulations showed that by carefully balancing how much each device trains, how much data they compress to save space, and how much they are trusted based on their current score, the group learns significantly better than if they just followed the old rules.

The researchers also discovered some interesting quirks about how these devices behave under pressure. They found that if the devices have to compress their data too much to save bandwidth (using fewer "quantization levels"), they might actually train more locally because they have more time and energy left over. However, if the AI model is too big and bulky, like the ResNet-18, increasing the compression can actually hurt performance because the devices run out of time to train at all. The paper suggests that there is no "one size fits all" setting; the best approach depends heavily on how powerful the device is and how much time it has.

Ultimately, this paper proposes a smarter, more flexible way for machines to learn together in a world where nothing stays the same. It argues that we can't just treat devices like static computers; we have to treat them like busy, resource-hungry partners that need a dynamic strategy to stay in sync. While these results are currently based on simulations and not yet tested on real-world hardware in a live network, the math and the computer models strongly suggest that this "score-aided" approach is a powerful tool for the future of privacy-friendly, efficient artificial intelligence.

Technical Summary: Online-Score-Aided Federated Learning for Resource-Constrained Wireless Clients with Continual Data Arrival

Problem Statement

Federated Learning (FL) in wireless networks faces significant challenges when deployed on resource-constrained clients (e.g., IoT devices) where data is not static. Traditional FL algorithms typically assume clients possess static datasets prior to training. However, in practical wireless applications such as integrated sensing and communication (ISAC), channel state information (CSI) prediction, and wearable sensors, training samples arrive continually.

This continual data arrival creates two critical, often overlooked issues:

Storage Constraints: Clients have limited storage capacity, shared with the operating system and user files. To accommodate new arriving samples, clients must delete old samples, leading to permanent data loss and potential "concept drift" where the optimal local model for the current dataset differs from the previous one.
Dynamic Distributions: Unlike traditional Continual Federated Learning (CFL) where new tasks or classes arrive, wireless applications often involve the same task (e.g., beam management) but with time-varying data distributions due to changing environmental conditions and the removal of old samples.

Existing FL algorithms (e.g., FedAvg, FedProx) primarily address system heterogeneity and non-IID data under static assumptions. They often perform poorly in these dynamic, resource-constrained settings because they do not account for the specific interplay between time-varying datasets, limited local storage, and the resulting shifts in data distribution.

Methodology: Online-Score-Aided Federated Learning (OSAFL)

The authors propose OSAFL, a new FL algorithm designed specifically for wireless clients with continual data arrival and strict resource constraints. The methodology integrates three core components:

1. Joint Resource Optimization

Before each training round, clients solve a non-convex optimization problem to determine their local training parameters. Given a fixed deadline ( $t_{th}$ ) and energy budget ( $e_{u,bd}$ ), clients jointly optimize:

The number of local stochastic gradient descent (SGD) rounds ( $\kappa_u^t$ ).
The CPU clock frequency ( $\bar{f}_u^t$ ).
The uplink transmission power ( $p_u^t$ ).

This ensures that clients maximize their local training effort while adhering to energy and time constraints, effectively managing the trade-off between computation and communication overheads.

2. Quantized Gradient Transmission

To reduce communication overhead, clients do not transmit full model updates. Instead, they compute normalized accumulated gradients over their local rounds and apply a stochastic quantizer. This process reduces the payload size significantly, allowing transmission within the limited wireless bandwidth and time slots.

3. Online Score-Aided Aggregation

The central server (CS) employs a novel aggregation policy that differs from standard weighted averaging. The global model update is governed by:
$w^{t+1} = w^t - \eta_{gl} \eta_{ef}^t \sum_{u=0}^{U-1} \alpha_u \delta_u^t \cdot \mathbb{1}_u^t \cdot Q(d_u^t)$

Key elements of this policy include:

Online Score ( $\delta_u^t$ ): A dynamic weight assigned to each client, calculated based on their participation frequency, local training rounds, and the cosine similarity between their quantized gradient and the global gradient direction. This score adapts to data distribution shifts and client reliability.
Global Step Size Controller ( $\eta_{ef}^t$ ): A scheduler that adjusts the effective step size based on the aggregated weights and the similarity of global gradients between consecutive rounds, aiming to minimize the convergence bound.
Participation Indicator ( $\mathbb{1}_u^t$ ): A binary variable determined by the resource optimization step; if a client cannot meet the deadline or energy constraints, they are excluded from that round.

Theoretical Analysis

The paper provides a rigorous convergence analysis of OSAFL. The authors derive an upper bound on the expected average global gradient norm, demonstrating how the convergence rate is affected by:

Data Distribution Shifts ( $\Phi^t$ ): The change in the global loss function due to continual data arrival and sample deletion.
Client Participation Uncertainty: The probability of clients being available ( $v_u^t$ ).
Quantization Noise: Errors introduced by the stochastic quantizer.
Gradient Dissimilarity: The divergence between local and global gradients caused by non-IID data.

The analysis reveals that for the algorithm to converge, the accumulated drift from data distribution shifts must not grow faster than $\sqrt{UT}$ (where $U$ is the number of clients and $T$ is the number of rounds). The derived bounds guide the selection of the online scores and step size controllers to minimize error terms.

Key Contributions

Algorithm Design: The proposal of OSAFL, the first FL algorithm explicitly designed for wireless clients with continual data arrival and strict storage limits, where old samples are deleted to make room for new ones.
Theoretical Framework: A convergence analysis that quantifies the impact of time-varying data distributions, uncertain participation, and gradient quantization on FL performance, leading to the derivation of optimal aggregation weights.
Sub-optimal Optimization: A practical method for calculating client scores and step sizes using gradient similarities and participation history, as exact optimization requires unknown future information.
Empirical Validation: Extensive simulations across three datasets (CIFAR10, FashionMNIST, MNIST) and three model architectures (SqueezeNet1, CNN, ResNet-18) with varying parameter counts.

Experimental Results

The authors evaluated OSAFL against modified versions of state-of-the-art baselines (M-FedAvg, M-FedProx, M-FedNova, M-FedDisco) and a Genie-aided centralized SGD baseline.

Performance: OSAFL consistently outperformed all modified baselines in terms of test accuracy and convergence speed across all datasets and models. For instance, on the CIFAR10 dataset with SqueezeNet1, OSAFL achieved test accuracies 4.5%–4.5% higher than the baselines at round 19.
Impact of Quantization: The study showed that the effect of quantization levels ( $s$ ) is model-dependent. For shallow models (SqueezeNet1), performance was relatively insensitive to $s$ . For moderate models (CNN), higher $s$ improved accuracy. However, for bulky models (ResNet-18), increasing $s$ increased payload size, leading to more stragglers and reduced performance under tight deadlines.
Impact of Heterogeneity: As the data heterogeneity parameter ( $\aleph$ ) increased (indicating more balanced label distributions), all algorithms improved, but OSAFL maintained a consistent lead.
Resource Constraints: The results confirmed that joint resource optimization is critical; without it, clients often fail to participate or complete sufficient local rounds, degrading global performance.

Significance and Claims

The paper claims that OSAFL addresses a critical gap in FL literature: the lack of algorithms capable of handling continual data arrival combined with strict storage constraints in wireless networks.

The authors emphasize that their work is distinct from traditional CFL, which focuses on new tasks, by focusing on the shifting distribution of the same task due to the necessity of deleting old data. They assert that their theoretical analysis provides necessary insights into how to weight client contributions in dynamic environments, and their simulations validate that sub-optimally minimizing convergence errors through online scoring leads to superior practical performance compared to existing methods adapted for this scenario. The work underscores that in resource-constrained wireless FL, the aggregation policy must be adaptive to data drift and client availability, rather than relying on static assumptions.