Abstract
Federated Domain Generalization (FedDG) aims to train a global model capable of generalizing to unseen domains using multi-domain training samples. However, in many real-world federated learning networks, clients are confined to a single, non-IID domain due to sampling and temporal limitations. This lack of cross-domain interaction and in-domain divergence—the single-source FedDG (sFedDG) problem—impedes the learning of domain-common features. To address this, we introduce Federated Global Consistent Augmentation (FedGCA). FedGCA incorporates a style-complement module to augment samples with diverse domain styles and employs global guided semantic and class consistency to mitigate local semantic noise and class imbalances. Extensive experiments demonstrate that FedGCA significantly outperforms existing state-of-the-art methods.
The Single-Source Problem in Federated Learning
Standard Federated Learning (FL) typically assumes that training and testing data are drawn from the same distribution. Federated Domain Generalization (FedDG) relaxes this assumption, aiming for robustness on "out-of-distribution" (OOD) data. However, most FedDG research assumes that the decentralized clients collectively cover multiple domains (e.g., some clients have photos, while others have sketches).
This paper identifies a more realistic but significantly harder challenge: single-source FedDG (sFedDG). In this scenario, all clients belong to just one domain, such as "Photo." This introduces two primary hurdles:
- Limited Style Diversity: Clients cannot naturally learn domain-invariant features because they only observe a single style.
- Semantic Inconsistency: Because FL data is non-IID, different clients possess varying class balances. Applying heavy data augmentation to synthesize new domains can inadvertently create noisy labels or inconsistent decision boundaries across the network.
Methodology: FedGCA
The proposed FedGCA (Federated Global Consistent Augmentation) framework addresses these challenges through three integrated components:
Style-Complement Module
To compensate for the lack of domain diversity, this module utilizes techniques such as RandConv (random convolutions) and AugMix to synthesize "pseudo-novel" versions of local images. These samples mimic different textures and styles while preserving the original object's identity, effectively simulating multi-domain data within a single-source environment.
Global Guided Semantic Consistency (GGSC)
Because synthesized data can be noisy, local clients use the Global Model as a teacher to maintain semantic integrity. Consistency is enforced at two levels:
- Coarse-grained: Aligning the softmax probability distributions of original and augmented images.
- Fine-grained: Utilizing Class Activation Maps (CAMs) to ensure the model focuses on the same physical regions of an image (e.g., the structural features of a guitar) regardless of its stylistic representation.
Class Consistency (CC)
To mitigate the drift caused by non-IID class distributions, FedGCA employs a dynamic regularizer. This aligns local model updates with the global objective, preventing individual clients from overfitting to their specific local class imbalances.
Experimental Results
The researchers evaluated FedGCA on two major domain generalization benchmarks: Digits (MNIST, SVHN, USPS, Synth, and MNIST-M) and PACS (Photo, Art, Cartoon, and Sketch).
Quantitative Performance
FedGCA demonstrated substantial leads over state-of-the-art methods like FedSAM, FedDyn, and FedSR:
- Digits Dataset: FedGCA achieved an average accuracy of 67.87%, outperforming the second-best method (FedSAM+RC) by 5.91%.
- PACS Dataset: On the challenging "Sketch" domain (trained only on Photos), FedGCA reached 41.19% accuracy, compared to just 25.18% for standard FedAvg.
Key Insights
- The Global Model as an Anchor: In sFedDG, local clients are prone to overfitting to noisy augmentations. By using the global model to guide consistency, the system leverages the collective knowledge of the network to filter out augmentation noise.
- CAMs for Generalization: Forcing the model to maintain consistent "attention" across styles helps it ignore stylistic background noise and focus on core object geometry.
- Augmentation Synergy: Results show that while data augmentation helps, it is insufficient on its own. The consistency losses are critical for translating augmented data into generalized performance.
Conclusion
FedGCA provides a robust framework for the single-source federated domain generalization problem. By combining sophisticated style augmentation with global-guided consistency, it enables decentralized networks to learn features that generalize far beyond their limited local distributions. Future work may focus on reducing the computational overhead of CAM calculations and extending the framework to complex tasks like object detection.
Summary updated