DeFed-GMM-DaDiL: A Decentralized Federated Framework for Domain Adaptation

TL;DR

DeFed-GMM-DaDiL is a decentralized federated framework that enables multi-source domain adaptation by modeling datasets as GMMs and sharing learnable components via Wasserstein barycenters, without a central server.

Contribution

It extends the GMM-DaDiL framework to a fully decentralized setting, allowing clients to collaboratively adapt to a target domain while preserving privacy.

Findings

Maintains stable shared representations across clients.

Effectively reconstructs missing classes in the target domain.

Achieves competitive results on domain adaptation benchmarks.

Abstract

Decentralized multi-source domain adaptation seeks to transfer knowledge from multiple heterogeneous and related source domains to an unlabeled target domain in a decentralized setting. We address this challenge through a fully decentralized federated approach, DeFed-GMM-DaDiL, an extension of the GMM-Dataset Dictionary Learning (DaDiL) framework. Each client models its dataset as a Gaussian Mixture Model (GMM), and the federation jointly approximates them via labeled Wasserstein barycenters of shared, learnable GMM atoms. This design enables adaptation without a central server while preserving clients' privacy. We empirically study the stability of the learned representations in scenarios where the target domain has missing classes. Empirical results demonstrate that DeFed-GMM-DaDiL maintains stable and consistent shared representations across clients, effectively reconstructs missing classes, and achieves competitive performance on multi-source domain adaptation benchmarks.