Seeking Global Flat Minima in Federated Domain Generalization via Constrained Adversarial Augmentation
Keywords: Federated Domain Generalization, Flat Minima, Data Augmentation
Abstract: Federated domain generalization (FedDG) aims at equipping the federally trained model with the domain generalization ability when the model meets new clients with domain shifts. Among factors that possibly indicate generalization, the loss landscape flatness of the trained model is an intuitive, viable, and widely studied one. However, pursuing the flatness of the global model in the FedDG setting is not trivial due to the restriction to preserve data privacy. To address this issue, we propose GFM, a novel algorithm designed to seek Global Flat Minima of the global model. Specifically, GFM leverages a global model-constrained adversarial data augmentation strategy, creating a surrogate for global data within each local client, which allows for split sharpness-aware minimization to approach global flat minima. GFM is compatible with federated learning without compromising data privacy restrictions, and theoretical analysis further supports its rationality by demonstrating that the objective of GFM serves as an upper bound on the robust risk of the global model on global data distribution. Extensive experiments on multiple FedDG benchmarks demonstrate that GFM consistently outperforms previous FedDG and federated learning approaches.
Supplementary Material: zip
Primary Area: applications to computer vision, audio, language, and other modalities
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10017
Loading