Efficient federated learning via aggregation of base models.

Area of Science:

• Distributed computing
• Machine learning
• Artificial intelligence

Background:

• Federated Learning (FL) offers enhanced privacy for training machine learning (ML) models.
• Standard FL client selection is random, effective for Independent and Identically Distributed (IID) data.
• Non-Independent and Identically Distributed (Non-IID) data in real-world scenarios degrades FL performance, causing lower accuracy and slower convergence.

Purpose of the Study:

• To address the performance degradation of FL in Non-IID settings.
• To propose a novel approach using base models and evolutionary algorithms for improved client selection.
• To enhance the accuracy and convergence speed of FL global models.

Main Methods:

• Proposed the concept of 'base models' representing diverse client data distributions.
• Demonstrated the theoretical existence of these base models.
• Employed evolutionary algorithms (EA) for optimizing client selection by encoding client IDs and utilizing operations like crossover and mutation.
• Integrated the EA-based client selection into existing FL frameworks (FedAvg, FedProx, SCAFFOLD).

Main Results:

• The proposed method significantly improves performance compared to random selection in FL.
• Faster convergence rates were observed across different FL frameworks and datasets.
• Experimental validation on FashionMNIST, MNIST, and TodayNews datasets confirmed superior results.

Conclusions:

• Base models effectively approximate diverse client distributions in FL.
• Evolutionary algorithms provide an efficient and effective alternative to random client selection for Non-IID data.
• The proposed approach enhances the practicality and efficiency of Federated Learning in real-world, heterogeneous data environments.