Improving the Efficiency of Federated Learning in Resource Allocation for Heterogeneous Networks through Intentional Client Selection

Arefi Jamal, Pouria; Safaei, Bardia

Please enable javascript in your browser.

Friend's email
Your name
Your email
enter code

This page was sent successfuly

Digital Object List

Bookmark

Review List

2406 viewed

بهبود کارایی یادگیری فدرال در تخصیص منابع شبکه‌های ناهمگون با انتخاب هدفمند کاربران

عارفی جمال، پوریا Arefi Jamal, Pouria

Improving the Efficiency of Federated Learning in Resource Allocation for Heterogeneous Networks through Intentional Client Selection

Arefi Jamal, Pouria | 2026

2 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 58898 (19)
University: Sharif University of Technology
Department: Computer Engineering
Advisor(s): Safaei, Bardia
Abstract:
The rapid expansion of the Internet of Medical Things (IoMT) and the massive generation of multimodal health data have provided unprecedented opportunities for early disease diagnosis and smart patient monitoring. However, the effective utilization of these data faces fundamental challenges regarding privacy preservation and severe computational resource constraints on edge devices. The distributed approach of Federated Learning (FL) addresses privacy concerns by training models locally without transferring raw data. Nevertheless, its implementation in real-world environments suffers from performance degradation due to device heterogeneity (the straggler effect) and their inability to run advanced, heavy learning models. To overcome these challenges, this thesis proposes a novel framework based on Split Federated Learning (SFL) for analyzing multimodal medical data. This approach leverages two key mechanisms: First, a smart client selection algorithm that evaluates nodes based on the ratio of “data utility to predicted training time,” effectively neutralizing the impact of straggler nodes. Second, a split architecture design wherein clients only utilize lightweight feature extractors and send the compressed features to the server. On the server side, Large Language Models (LLMs) are employed to understand the complex relationships within the multimodal data. To reduce computational costs, Low-Rank Adaptation (LoRA), a parameter-efficient fine-tuning technique, is implemented. The results demonstrate that the proposed framework not only guarantees the privacy of sensitive patient data but also significantly reduces network latency and communication costs by offloading processing to the server and optimally selecting clients, thereby enabling the deployment of state-of-the-art AI models in resource-constrained IoT hardware environments
Keywords:
Internet of Medical Things ; Multi-Modal Data ; Low-Rank Adaptation (LoRA) ; Large Language Model ; User Admission ; Split Federated Learning (SFL)

Digital Object List

محتواي کتاب
view

Bookmark