DFPL: Decentralized Federated Prototype Learning Across Heterogeneous Data Distributions

2025-05-08

AI Summary

To address performance degradation caused by statistical heterogeneity in decentralized federated learning, this paper proposes the Decentralized Federated Prototype Learning (DFPL) framework. DFPL is the first to integrate prototype-based learning into a serverless distributed training paradigm, mitigating data distribution shift via local prototype modeling and enabling joint scheduling of communication, computation, and consensus resources through a co-designed training-and-blockchain-mining mechanism. We provide theoretical convergence guarantees and introduce a resource-aware distributed optimization analysis method. Experiments on three heterogeneous datasets demonstrate that DFPL achieves an average 4.2% improvement in test accuracy, reduces communication overhead by 67%, and accelerates convergence by 2.1×, significantly enhancing both model robustness and system efficiency.

Technology Category

• Machine Learning: Distributed Machine Learning & Federated Learning
• Search and Optimization: Distributed Search
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Systems and Infrastructure for Web, Mobile and WoT: Decentralized Web and Fediverse systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Federated learning is a distributed machine learning paradigm that enables the collaborative training of multiple clients through centralized model aggregation. However, standard federated learning relies on a centralized server, making it vulnerable to server failures. While existing solutions utilize blockchain technology to implement Decentralized Federated Learning (DFL), the statistical heterogeneity of data distributions among clients severely degrades the DFL's performance. Driven by this issue, this paper proposes a decentralized federated prototype learning framework, named DFPL, which significantly improves the performance of distributed machine learning across heterogeneous data distributions. Specifically, our framework introduces prototype learning into DFL to address statistical heterogeneity, which greatly reduces the number of parameters exchanged between clients. Additionally, blockchain is embedded into our framework, enabling the training and mining processes to be implemented at each client. From a theoretical perspective, we provide convergence guarantee of DFPL by combining resource allocation for training and mining. The experiments highlight the superiority of our DFPL framework in communication efficiency and test performance across three benchmark datasets with heterogeneous data distributions.

Problem

• Decentralized federated learning vulnerable to data heterogeneity
• Reducing parameter exchange in heterogeneous client distributions
• Integrating blockchain for client-based training and mining

Innovation

• Decentralized federated prototype learning framework
• Prototype learning reduces exchanged parameters
• Blockchain enables client training and mining