FedDNA: Federated learning using dynamic node alignment.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Federated Learning (FL) offers privacy-preserving model training.
• Current FL methods often use static node alignment, which can be suboptimal.
• The roles of individual nodes in neural networks are complex and dynamic.

Purpose of the Study:

• To introduce FedDNA, a novel federated learning algorithm utilizing dynamic node alignment.
• To enhance the performance of federated learning by improving node matching across distributed sites.
• To address the limitations of static node matching in existing federated learning approaches.

Main Methods:

• Representing neural network node weights as vectors.
• Employing distance functions to identify similar nodes across distributed sites.
• Utilizing a minimum spanning tree approach for efficient and comprehensive node matching.

Main Results:

• FedDNA dynamically aligns nodes for improved federated learning.
• The algorithm effectively finds optimal node matches across different sites.
• Experimental results show FedDNA surpasses traditional methods like FedAvg.

Conclusions:

• Dynamic node alignment is a promising strategy for advancing federated learning.
• FedDNA offers a computationally efficient and effective solution for node matching.
• The proposed method enhances model performance and privacy in federated learning settings.