MEFL: Meta-Equilibrize Federated Learning for Imbalanced Data in IoT.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Internet of Things

Background:

• Data heterogeneity in IoT environments poses significant challenges to Federated Learning (FL) models.
• Imbalanced and non-IID data distributions degrade model accuracy, robustness, and generalization capabilities.

Purpose of the Study:

• To propose a novel method, Meta-Equilibrized Federated Learning (MEFL), to address data heterogeneity and improve FL performance in IoT.
• To enhance the consistency between global and local optimization objectives in FL.

Main Methods:

• MEFL integrates meta-learning with gradient-descent preservation.
• It employs an equilibrated optimization aggregation mechanism using gradient similarity and variance-weighted adjustment.
• The method alleviates gradient biases from multi-step local updates.

Main Results:

• MEFL achieved at least a 3.26% improvement in final test accuracy compared to baseline methods.
• The approach substantially reduced communication overhead.
• Demonstrated superior performance and generalization capabilities on real-world datasets.

Conclusions:

• MEFL effectively resolves inconsistencies between global and local optimization objectives in FL.
• The method optimizes trade-offs between local and global models for personalized IoT applications.
• MEFL offers an efficient solution for cross-domain data security deployment in IoT.