Joint multi-dimensional resource optimization for model compression in wireless federated learning
In the edge computing scenarios, resource-constrained and particiption of the dynamically terminal devices of network in federated learning cause high latency and high energy consumption. An efficient and environmentally friendly federated learning algorithm based on a three-tier cloud-edge-terminal...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | Chinese |
| Published: |
China InfoCom Media Group
2025-06-01
|
| Series: | 物联网学报 |
| Subjects: | |
| Online Access: | http://www.wlwxb.com.cn/zh/article/doi/10.11959/j.issn.2096-3750.2025.00391/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In the edge computing scenarios, resource-constrained and particiption of the dynamically terminal devices of network in federated learning cause high latency and high energy consumption. An efficient and environmentally friendly federated learning algorithm based on a three-tier cloud-edge-terminal architecture was proposed. Firstly, by introducing model compression techniques into the three-tier federated learning structure, a theoretical analysis was conducted on the model convergence rate, training latency, and energy consumption. Subsequently, based on the theoretical analysis, a problem was formulated to minimize the global model training latency and energy consumption under a certain model convergence rate by jointly optimizing the terminal devices' transmission power, computing power, and model compression rate. Finally, by decomposing the problem into three sub-optimization problems and solving them alternately, a joint alternating optimization algorithm was designed to obtain the optimal solution for the original problem. Experimental results demonstrate that the proposed algorithm is adaptable to large-scale edge computing scenarios. It achieves reductions of 71.54% and 48.76%, respectively, in latency and energy consumption compared with traditional three-layer federated learning algorithms, while ensuring the convergence rate of the model, and effectively reduces the latency and energy consumption generated by global model training. |
|---|---|
| ISSN: | 2096-3750 |