FedEmerge: An Entropy-Guided Federated Learning Method for Sensor Networks and Edge Intelligence

Area of Science:

• Distributed machine learning
• Edge intelligence
• Data privacy

Background:

• Federated Learning (FL) enables collaborative model training without raw data sharing, crucial for privacy and bandwidth in sensor networks.
• Traditional FL methods like FedAvg struggle with non-IID data, common in edge environments.
• Existing aggregation methods may overlook data informativeness, impacting performance on sparse or unbalanced client datasets.

Purpose of the Study:

• To introduce FedEmerge, an entropy-guided aggregation approach for Federated Learning.
• To enhance global model performance by prioritizing client updates based on data diversity and information entropy.
• To enable emergent collective learning dynamics for improved convergence in heterogeneous networks.

Main Methods:

• Developed FedEmerge, an entropy-guided aggregation strategy for FL.
• Quantified client data diversity using information entropy to adjust update weights.
• Proved FedEmerge's convergence under the Polyak-Łojasiewicz (PL) condition.

Main Results:

• FedEmerge achieves linear convergence, comparable to centralized gradient descent.
• Empirical results show improved accuracy and faster convergence on skewed non-IID benchmarks.
• Reduced performance disparities among clients compared to FedAvg across CIFAR-10, Federated EMNIST, and Shakespeare datasets.

Conclusions:

• Entropy-guided aggregation enhances FL outcomes in heterogeneous networks by weighting client updates by data diversity.
• FedEmerge offers a practical, privacy-preserving solution with minimal computational overhead for real-world federated systems.