Federated k-means based on clusters backbone

Area of Science:

• Machine Learning
• Distributed Systems
• Data Mining

Background:

• Federated clustering is a privacy-preserving distributed algorithm.
• Non-Independent and Identically Distributed (Non-IID) data presents challenges for global consistency in federated learning.
• Existing federated clustering methods often exhibit suboptimal performance on Non-IID datasets.

Purpose of the Study:

• To propose a novel federated k-means clustering algorithm, FKmeansCB, designed to effectively handle Non-IID data.
• To improve the accuracy and efficiency of federated clustering in distributed environments.

Main Methods:

• Developed FKmeansCB, a federated k-means algorithm utilizing a cluster backbone.
• Implemented Laplace noise addition to local data for robust cluster center representation.
• Designed a server-client aggregation strategy for global cluster center computation.

Main Results:

• FKmeansCB demonstrated significant improvements in clustering accuracy compared to existing methods.
• The algorithm showed substantial reductions in running time across multiple datasets.
• Validated performance on large-scale datasets, including MNIST, under Non-IID conditions.

Conclusions:

• FKmeansCB effectively addresses the challenge of Non-IID data in federated clustering.
• The cluster backbone approach enhances the representation of local data structures.
• FKmeansCB offers a promising solution for accurate and efficient distributed clustering.