Client and Training Data Selection for Computationally Efficient Synchronized Federated Learning

The paper Client and Training Data Selection for Computationally Efficient Synchronized Federated Learning (arXiv:2609.39250v1) proposes a joint client and training data selection algorithm to address straggling clients and non-IID data in Federated Learning. By balancing training dataset size, class distribution, and data staleness, the method ensures fast convergence under strict timing constraints.

Key contributions and findings include: Probabilistic Selection: The algorithm uses a probabilistic approach to estimate training times and task interruptions, maximizing the expected contribution of training data per round. Non-IID Awareness: It explicitly balances class distribution during client selection, outperforming baselines like MinCost and probPart which struggle with uneven data distributions. Performance Gains: Experiments on the CIFAR-100 dataset demonstrated that the proposed method achieves significantly improved convergence and higher model accuracy compared to prior works, particularly in scenarios with heterogeneous computation and small client pools. Deterministic Efficiency: Unlike probabilistic prior approaches, this method deterministically selects the most helpful clients, preventing waste of computation resources on devices that do not participate.

This paper studies synchronized federated learning (FL), where a central server updates the global model only after receiving local updates from a participating set of clients. The central problem it addresses is the impact of straggling clients: slow devices or computationally expensive local training steps delay aggregation, increase round latency, and degrade the convergence speed of the global model. The authors treat this not merely as a systems scheduling issue, but as a joint optimization problem involving which clients should participate and how much local training data each selected client should use in a given round. By coordinating client selection with local data selection, the work aims to reduce the aggregate computation and waiting time required per synchronization round while preserving enough useful gradient information for effective global learning.

A key insight is that efficient FL under stragglers requires balancing computation cost, client heterogeneity, and data informativeness. Simply excluding slow clients may remove useful or representative data, while including all clients can make the system bottlenecked by the slowest participants. Likewise, reducing local training data can lower compute and communication overhead, but overly aggressive subsampling may weaken the local update and harm convergence. The paper therefore frames client and data selection as a coupled trade-off: choose a subset of clients and a subset of their local data so that the server receives sufficiently high-quality updates in a bounded time, improving computational efficiency without sacrificing model performance.

This work matters because synchronous FL is a common deployment model, but real-world federated systems run on heterogeneous edge devices with limited battery, storage, bandwidth, and compute. Straggler-induced delays can make FL impractical or inefficient in large-scale settings, especially under non-IID data distributions. By improving how clients and local data are selected, the paper contributes to making FL more scalable, energy-efficient, and latency-aware. It is relevant to researchers and practitioners working on distributed optimization, edge AI, privacy-preserving learning, and systems where convergence quality, resource constraints, and synchronization overhead must be jointly managed.