| Title |
Clustering-based Decentralized Federated Learning Framework with Aggregator Nodes |
| Authors |
박준영(Junyoung Park) ; 이주형(Joohyung Lee) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.8.15 |
| Keywords |
Decentralized federated learning; Aggregator node; Personalization |
| Abstract |
Decentralized federated learning(DFL) is a distributed collaborative learning approach where nodes exchange and aggregate models directly. It has the advantage of overcoming single points of failure and bottlenecks in central servers compared to conventional federated learning. However, the absence of an aggregator leads to performance degradation compared to conventional federated learning. Additionally, in distributed learning, it is important to make the model have both generalization and personalization capabilities. This study proposes a clustering-based decentralized federated learning framework to overcome the performance degradation caused by the absence of an aggregator node while ensuring both personalization and generalization capabilities. The proposed method in this paper conducts clustering based on optimization path similarity to effectively introduce aggregators. Intra-cluster learning aims to achieve personalization capability by utilizing aggregator nodes to collaborate with nodes that have similar learning directions. Subsequently, to prevent excessive personalization and ensure generalization capability, inter-cluster learning is conducted. Experimental analysis on the efficacy of aggregator in decentralized federated learning was conducted, demonstrating a 4.2% improvement in accuracy compared to decentralized federated learning on the CIFAR-10 dataset. |