Work

FedSimRep: A Federated Learning Method Based on Data Representation Selection

Year: 2025

Type: conference-paper

Abstract: As an emerging distributed technology, Federated learning enables model training while protecting privacy without sharing local data. However, the heterogeneity of geographic locations and devices leads to response delays from stragglers. Client grouping in federated learning can tolerate stragglers from a global perspective, but differences in data distribution within groups may result in local performance bias. To address this issue, we propose a federated learning method called FedSimRep, which selects representative clients based on data distribution similarity. FedSimRep performs group training by collecting readiness signals from topological communications and maintains synchronization tables and graphs to track training progress. Additionally, it selects representative clients by leveraging the relationship between local optimization similarity and data distribution similarity to optimize the global model. Experiments and theoretical analyses on public datasets such as MNIST, CIFAR-10, and UEU-Fed demonstrate that FedSimRep achieves significant advantages in terms of accuracy and efficiency.

Source: International Conference on Communication Networks and Machine Learning (CNML)

DOI: 10.1145/3728199.3728206

Authors: Falu Zhai, Lu Wu, Chunxiao Wang, Tongyang Hao, Shucang Zang

Institutions: Qilu University of Technology, Shandong Academy of Sciences, National Supercomputing Center in Jinan

Language: English

Topic: Privacy-Preserving Technologies in Data

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Funders: Natural Science Foundation of Shandong Province, National Supercomputing Center, Korea Institute of Science and Technology Information

Award: G1479536537