Multi-Source Collaborative Gradient Discrepancy Minimization for Federated Domain Generalization

TL;DR

This work tackles the challenge of learning domain-invariant models under privacy-preserving federated constraints. It introduces Multi-source Collaborative Gradient Discrepancy Minimization (MCGDM), which combines intra-domain gradient matching between original and augmented images with inter-domain gradient matching across decentralized domains to mitigate both within-domain and cross-domain shifts. The method operates within a FedDG framework and extends to Federated Domain Adaptation by leveraging pseudo-labels for the target domain, achieving state-of-the-art results on several benchmarks. Empirical results demonstrate robust improvements over existing FedDG and conventional DG methods, with clear ablations validating the effectiveness of gradient matching and augmentation strategies. This approach offers a privacy-preserving, scalable path to domain generalization and adaptation in distributed settings.

Abstract

Federated Domain Generalization aims to learn a domain-invariant model from multiple decentralized source domains for deployment on unseen target domain. Due to privacy concerns, the data from different source domains are kept isolated, which poses challenges in bridging the domain gap. To address this issue, we propose a Multi-source Collaborative Gradient Discrepancy Minimization (MCGDM) method for federated domain generalization. Specifically, we propose intra-domain gradient matching between the original images and augmented images to avoid overfitting the domain-specific information within isolated domains. Additionally, we propose inter-domain gradient matching with the collaboration of other domains, which can further reduce the domain shift across decentralized domains. Combining intra-domain and inter-domain gradient matching, our method enables the learned model to generalize well on unseen domains. Furthermore, our method can be extended to the federated domain adaptation task by fine-tuning the target model on the pseudo-labeled target domain. The extensive experiments on federated domain generalization and adaptation indicate that our method outperforms the state-of-the-art methods significantly.

Figures (3)

• Figure 1: (a) DG assumes that the data from multiple source domains $\{X_{i},Y_{i}\}_{i=1}^{n}$ can be accessed simultaneously to learn a generalized model $\mathcal{M}$ for deployment on the unseen domain $X_{T}$. (b) FedDG assumes that the data from different source domains are decentralized, but the local models $\{\mathcal{M}_{i}\}_{i=1}^{n}$ of different domains can be collaboratively trained and aggregated with a parameter server. (c) FedDA assumes that an additional unlabeled target domain $X_{T}$ can be accessed on server side for improving the performance.
• Figure 2: (a) The overall framework of our method, where the local source domain models $\{\mathcal{M}_{i}\}_{i=1}^{n}$ are trained locally by conducting gradient matching and then are aggregated on the server side to obtain the generalizable global model $\mathcal{M}_{G}$. (b) The intra-domain gradient matching between the original images and the augmented images is conducted on local clients for learning the intrinsic semantic information within the domain. (c) The inter-domain gradient matching between the current classifier head $\mathcal{C}_{i}$ and the classifier heads from other domains $\{\mathcal{C}_{j}^{t-1}\}_{j=1}^{n}$ is conducted on the local clients for reducing the domain shift across decentralized source domains.
• Figure 3: Visualization by Grad-CAM on unseen domain.