Decentralized Federated Domain Generalization with Style Sharing: A Formal Modeling and Convergence Analysis

2025-04-08

AI Summary

This paper addresses domain generalization (DG) in decentralized federated learning (FL), noting that existing work lacks formal modeling of DG objectives and training dynamics, and is constrained by star-topology architectures, thus failing to leverage the full potential of decentralized networks. To this end, we propose StyleDDG: a serverless DG algorithm for peer-to-peer networks that achieves domain generalization via style-feature disentanglement and cross-device style sharing. We establish the first formal framework for style-driven DG, including a rigorous sublinear convergence analysis. Furthermore, we introduce a graph neural network–inspired collaborative update mechanism. Extensive experiments on two benchmark DG datasets demonstrate significant improvements in target-domain accuracy, while communication overhead remains comparable to standard decentralized gradient methods. Both theoretical analysis and empirical evaluation jointly validate the efficacy and convergence guarantees of StyleDDG.

Technology Category

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

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• User Modeling, Personalization and Recommendation: Federated recommendation systems and personalization
• Systems and Infrastructure for Web, Mobile and WoT: Decentralized Web and Fediverse systems

Abstract

Much of the federated learning (FL) literature focuses on settings where local dataset statistics remain the same between training and testing time. Recent advances in domain generalization (DG) aim to use data from source (training) domains to train a model that generalizes well to data from unseen target (testing) domains. In this paper, we are motivated by two major gaps in existing work on FL and DG: (1) the lack of formal mathematical analysis of DG objectives and training processes; and (2) DG research in FL being limited to the conventional star-topology architecture. Addressing the second gap, we develop $ extit{Decentralized Federated Domain Generalization with Style Sharing}$ ($ exttt{StyleDDG}$), a fully decentralized DG algorithm designed to allow devices in a peer-to-peer network to achieve DG based on sharing style information inferred from their datasets. Additionally, we fill the first gap by providing the first systematic approach to mathematically analyzing style-based DG training optimization. We cast existing centralized DG algorithms within our framework, and employ their formalisms to model $ exttt{StyleDDG}$. Based on this, we obtain analytical conditions under which a sub-linear convergence rate of $ exttt{StyleDDG}$ can be obtained. Through experiments on two popular DG datasets, we demonstrate that $ exttt{StyleDDG}$ can obtain significant improvements in accuracy across target domains with minimal added communication overhead compared to decentralized gradient methods that do not employ style sharing.

Problem

• Lack of formal analysis for domain generalization objectives
• Limited FL domain generalization to star-topology architectures
• Need decentralized DG via peer-to-peer style sharing

Innovation

• Decentralized peer-to-peer style sharing network
• Mathematical analysis of style-based DG training
• Sub-linear convergence with style sharing