Resource Allocation for Federated Learning with Heterogeneous Computing Capability in Cloud–Edge–Client IoT Architecture
A federated learning (FL) framework for cloud–edge–client collaboration performs local aggregation of model parameters through edges, reducing communication overhead from clients to the cloud. This framework is particularly suitable for Internet of Things (IoT)-based secure computing scenarios that...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-05-01
|
| Series: | Future Internet |
| Subjects: | |
| Online Access: | https://www.mdpi.com/1999-5903/17/6/243 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | A federated learning (FL) framework for cloud–edge–client collaboration performs local aggregation of model parameters through edges, reducing communication overhead from clients to the cloud. This framework is particularly suitable for Internet of Things (IoT)-based secure computing scenarios that require extensive computation and frequent parameter updates, as it leverages the distributed nature of IoT devices to enhance data privacy and reduce latency. To address the issue of high-computation-capability clients waiting due to varying computing capabilities under heterogeneous device conditions, this paper proposes an improved resource allocation scheme based on a three-layer FL framework. This scheme optimizes the communication parameter volume from clients to the edge by implementing a method based on random dropout and parameter completion before and after communication, ensuring that local models can be transmitted to the edge simultaneously, regardless of different computation times. This scheme effectively resolves the problem of high-computation-capability clients experiencing long waiting times. Additionally, it optimizes the similarity pairing method, the Shapley Value (SV) aggregation strategy, and the client selection method to better accommodate heterogeneous computing capabilities found in IoT environments. Experiments demonstrate that this improved scheme is more suitable for heterogeneous IoT client scenarios, reducing system latency and energy consumption while enhancing model performance. |
|---|---|
| ISSN: | 1999-5903 |