SemiDFL: A Semi-Supervised Paradigm for Decentralized Federated Learning
Xinyang Liu, Pengchao Han, Xuan Li, Bo Liu·December 18, 2024
Summary
SemiDFL is a semi-supervised decentralized federated learning method addressing non-IID scenarios. It enhances model training by establishing consensus in data and model spaces, utilizing neighborhood information for improved pseudo-labeling and designing a consensus-based diffusion model to generate synthesized data. Adaptive aggregation leverages synthesized data's model accuracy, outperforming existing centralized and decentralized federated learning schemes in both IID and non-IID SSL scenarios.
Introduction
Background
Overview of federated learning
Challenges in non-IID data distribution
Importance of semi-supervised learning in federated settings
Objective
To present SemiDFL as a solution for enhancing model training in non-IID federated learning environments
Highlighting the method's ability to improve pseudo-labeling and model consensus
Method
Data Collection
Explanation of decentralized data collection in federated learning
Importance of neighborhood information in data gathering
Data Preprocessing
Techniques for handling non-IID data in a semi-supervised context
Methods for preparing data for consensus-based diffusion
Pseudo-Labeling
Role of neighborhood information in enhancing pseudo-labeling accuracy
Process of generating reliable pseudo-labels for unlabeled data
Consensus-Based Diffusion Model
Design and implementation of the model for generating synthesized data
Explanation of how synthesized data improves model training
Adaptive Aggregation
Mechanism for leveraging synthesized data's model accuracy
Comparison with existing centralized and decentralized federated learning schemes
Evaluation
Performance Metrics
Metrics used to assess SemiDFL's effectiveness in both IID and non-IID SSL scenarios
Comparative Analysis
Detailed comparison of SemiDFL against existing methods
Discussion on performance improvements in non-IID federated learning environments
Conclusion
Summary of Contributions
Recap of SemiDFL's unique features and benefits
Future Work
Potential areas for further research and development
Implications for broader applications in decentralized federated learning
Basic info
papers
distributed, parallel, and cluster computing
machine learning
artificial intelligence
Advanced features