Multi-Source Collaborative Style Augmentation and Domain-Invariant Learning for Federated Domain Generalization
Yikang Wei
TL;DR
This paper introduces MCSAD, a novel federated domain generalization method that enhances style augmentation across multiple sources and learns domain-invariant features to improve generalization to unseen domains.
Contribution
The paper proposes a multi-source collaborative style augmentation module combined with domain-invariant learning for federated domain generalization, addressing style space limitations.
Findings
Significantly outperforms state-of-the-art federated domain generalization methods.
Effectively broadens style space through collaborative augmentation.
Improves generalization to unseen target domains.
Abstract
Federated domain generalization aims to learn a generalizable model from multiple decentralized source domains for deploying on the unseen target domain. The style augmentation methods have achieved great progress on domain generalization. However, the existing style augmentation methods either explore the data styles within isolated source domain or interpolate the style information across existing source domains under the data decentralization scenario, which leads to limited style space. To address this issue, we propose a Multi-source Collaborative Style Augmentation and Domain-invariant learning method (MCSAD) for federated domain generalization. Specifically, we propose a multi-source collaborative style augmentation module to generate data in the broader style space. Furthermore, we conduct domain-invariant learning between the original data and augmented data by cross-domain…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
1. **Originality:** The paper addresses a genuine problem in federated domain generalization, the limited style space when data cannot be shared across domains. Using remote classifier heads as discriminators avoids explicit sharing of style statistics or features, which is architecturally sensible for privacy-preserving learning. 2. **Experiments:** The experimental evaluation are reasonably comprehensive, covering three standard domain generalization benchmarks with appropriate train/test spli
1. **Lack of Privacy Analysis:** While the paper claims reduced privacy leakage compared to sharing features/full models (Appendix D), this analysis is superficial. This is significant given that privacy is presented as a key motivation and individuals looking to adopt these methods would benefit from a more thorough discussion or empirical validation of privacy guarantees. 2. **Limited Evaluation Scope and Realism:** The experiments are conducted on standard domain generalization benchmarks (PA
The paper is clearly written, well-structured, and highly readable.
1. A lack of comparison with the key approach: I noticed that the main evaluation table only compares with studies published before 2023, lacking comparisons with the more recent works from 2024 and 2025. The key comparison FDG method [1] was not even cited. 2. Similar to previous work: After a careful comparison with prior work, I find that the proposed framework is highly similar to existing FDG approach [1], lacking significant innovation. In particular, Figure 2(a) in this paper is entirely
* The motivation for combining collaborative augmentation and domain-invariant learning is well justified. * This problem is timely, given increasing interest in privacy-preserving domain generalization. * The framework (Fig. 2) and pseudo-code-like descriptions (Steps 1–4, Sec. 3.1) provide a clear understanding of the workflow. * The improvements are notable, particularly 86.3% on PACS and 78.8% on VLCS, surpassing centralized MixStyle and RSC.
* No ablation or visualization quantifies the degree of style diversity or the domain invariance learned (no measure in Sec. 4). * It remains unclear whether collaborative augmentation increases local computational burden. * The update rules for µ̂, σ̂ could benefit from clearer explanation or pseudo-code.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsPrivacy-Preserving Technologies in Data