LyRA-FL: A Scheduler for Federated Fine-Tuning of Foundation Models in Wireless Networks
Year: 2026
Type: article
Abstract: The distributed training of semantic segmentation models is a key enabler for perception on mobile and edge devices, supporting applications such as augmented reality, robotics, and intelligent video analytics. However, this deployment within federated learning (FL) systems is hindered by severe statistical heterogeneity, high communication costs, and the unreliability of wireless networks. This paper introducesLyapunov-basedResourceAllocation for FL (LyRA-FL), a unified cross-layer scheduling algorithm grounded in control theory and stochastic geometry. At the learning layer, we integrate the SegFormer foundation vision model with Low-Rank Adaptation (LoRA) to enable parameter-efficient fine-tuning and reduce the communication and computation burden on clients. At the communication layer, LyRA-FL employs a stochastic geometry-driven Lyapunov scheduler that jointly optimizes communication reliability, a novel Value-of-Participation (VoP) metric, and long-term fairness during client selection. Extensive experiments on the Cityscapes and IDDA benchmarks show that our unified framework achieves over 82% reduction in the number of trainable parameters and converges significantly faster than state-of-the-art baselines, achieving up to 10% and 33% higher final mean Intersection-over-Union (mIoU), respectively. Our LyRA-FL scheduler successfully manages issues like wireless uncertainty, Non-Independent and Identically Distributed (non-IID) data, and stragglers, enabling our framework to achieve a higher final accuracy under realistic channel conditions than the SiloBN+LAB baseline in an idealized, error-free setting.
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7203609236