• Multi-Source Collaborative Gradient Discrepancy Minimization for Federated Domain Generalization
• • Abstract
  • Different setups of domain generalization
  • Method
  • • Install Datasets
    • FedDG
  • Reference
  • Acknowledgments

Multi-Source Collaborative Gradient Discrepancy Minimization for Federated Domain Generalization

Here is the Mindspore Implementation of the paper ["Multi-Source Collaborative Gradient Discrepancy Minimization for Federated Domain Generalization", AAAI 2024].

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.

Different setups of domain generalization

• Domain Generalization (DG) assumes that the data from multiple source domains can be accessed simultaneously to learn a generalizable model for deployment on the unseen domain.
• Federated Domain Generalization (FedDG) assumes that the data from different source domains are decentralized, but the local models of different domains can be collaboratively trained and aggregated with a parameter server.
• Federated Domain Adaptation (FedDA) assumes that an additional unlabeled target domain can be accessed on server side for improving the performance.

Method

• (1) Collaborative train the decentralized domains: Due to the data from different domains are decentralized, we utilize the federated learning framework, e.g. FedAvg to collaborative train the multiple decentralized source domains.
• (2) Gradient discrepancy indicates the domain-specific information: Inspired by the hypothesis that the gradient discrepancy between domains indicates the model updating to be domain-specific, we utilize the gradient discrepancy to detect the domain shift.
• (3) Reducing the domain shift under the data decentralization scenario: Under the data decentralization scenario, we propose to learn the intrinsic semantic information within isolated domain and reduce the domain shift between decentralized domains by reducing the gradient discrepancy within domain and across domains.

Install Datasets

Please prepare the PACS dataset.

base_path
│       
└───dataset
│   │   pacs
│       │   images
│       │   splits

FedDG

The configuration files can be found under the folder ./config, and we provide four config files with the format .yaml. To perform the FedDG on the specific dataset (e.g., PACS), please use the following commands:

CUDA_VISIBLE_DEVICES=0 python main_dg.py --config PACS.yaml --target-domain art_painting -bp ../

CUDA_VISIBLE_DEVICES=0 python main_dg.py --config PACS.yaml --target-domain cartoon -bp ../

CUDA_VISIBLE_DEVICES=0 python main_dg.py --config PACS.yaml --target-domain photo -bp ../

CUDA_VISIBLE_DEVICES=0 python main_dg.py --config PACS.yaml --target-domain sketch -bp ../

Reference

If you find this useful in your work please consider citing:

@article{wei2024multi,
  title={Multi-Source Collaborative Gradient Discrepancy Minimization for Federated Domain Generalization},
  author={Wei, Yikang and Han, Yahong},
  journal={AAAI},
  year={2024}
}

Acknowledgments

This work is supported by the CAAI-Huawei MindSpore Open Fund.