Related Experiment Videos
Digital Twin-Enabled Dynamic Aggregation for Efficient Federated Learning
Wenqin Zhuang1, Yuao Wang1, Guocheng Wang1
1Jiangsu Key Laboratory of Intelligent Information Processing and Communication Technology, Nanjing University of Posts and Telecommunications, Nanjing 210003, China.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a digital twin (DT) approach for dynamic federated learning (FL) aggregation, enhancing efficiency and accuracy in heterogeneous networks. The DT optimizes client grouping and aggregation strategies, reducing latency and energy use.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Distributed Systems
Background:
- Federated learning (FL) allows collaborative training without raw data sharing.
- Client heterogeneity in FL leads to inefficiency and decreased model accuracy.
- Existing FL aggregation methods struggle with diverse client characteristics.
Purpose of the Study:
- To propose a novel digital twin (DT)-based dynamic aggregation method for federated learning.
- To address challenges posed by client heterogeneity in FL systems.
- To optimize FL performance by minimizing latency and energy consumption while maximizing accuracy.
Main Methods:
- Implementation of a digital twin (DT) layer for pre-aggregation evaluation and strategy simulation.
- Utilizing adaptive K-means clustering for grouping clients with similar characteristics.
- Designing a hierarchical aggregation evaluation strategy for intra-cluster and inter-cluster optimization.
Main Results:
- The proposed DT-assisted FL method accelerates model convergence and improves accuracy.
- Significant reductions in training latency and energy consumption were observed compared to baseline FL algorithms.
- Demonstrated effectiveness on MNIST and CIFAR-10 datasets.
Conclusions:
- The digital twin-based dynamic aggregation method offers a practical and effective solution for federated learning.
- This approach optimizes large-scale heterogeneous IoT sensor networks.
- DTs provide a powerful tool for enhancing federated learning performance and deployment.
