Image Style Transfer-Empowered Federated Domain Generalization

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Federated learning enables privacy-preserving collaborative model training.
• Domain shift due to data heterogeneity hinders federated learning generalization.
• Existing methods struggle with diverse, unseen target domains.

Purpose of the Study:

• To propose an image style transfer-empowered federated domain generalization method.
• To enhance the generalization performance of federated learning models.
• To address the challenge of domain shift in practical federated learning applications.

Main Methods:

• Enriching local data diversity using image style transfer.
• Implementing predictive consistency regularization for stable outputs and domain-invariant features.
• Developing a generalization capability-aware aggregation weight optimization strategy using simulated target domains.

Main Results:

• The proposed method achieves higher average test accuracy and faster convergence on PACS and Office-Home datasets.
• Experiments on the Camelyon17 dataset confirm the method's robustness and practical applicability.
• Validation of both local regularization and aggregation weight optimization strategies.

Conclusions:

• The image style transfer-empowered federated domain generalization method effectively improves model performance on unseen domains.
• The predictive consistency regularization and adaptive aggregation strategies enhance generalization capabilities.
• The approach demonstrates robustness and practical utility in real-world scenarios.