[4M3-GS-10] AI application: Optimization / Visualization

[4M3-GS-10-05] A Preliminary Study on Blockchain-based Clustered Federated Learning

〇Hiiro Uchiyama¹, Shota Suzuki¹, Satoshi Ono¹ (1. Kagoshima University)

[[Online]]

Keywords:Federated Learning, Clustering, Blockchain

Clustered Federated Learning categorizes participants into clusters to build models within each cluster, effectively dealing with heterogeneous environments where data distributions vary among participants. However, the current methods' limitation lies in their inflexibility to handle heterogeneity due to the predetermined number of clusters and the implementation of hard clustering. To address this, Federated Learning via Inference Similarity (FLIS) was proposed, utilizing inference similarity of participants' models for cluster formation. Despite its advantages, centralized federated learning approaches, including FLIS, face vulnerabilities from the dependency on a central server for managing data and learning processes. Consequently, we propose a method that implements FLIS learning on a blockchain network, allowing dynamic cluster formation based on participants' data characteristics without the need for a central server, thus supporting model training for each cluster. Our experiments show that this method achieves performance on par with FLIS while making clustered federated learning possible on a blockchain network.