Federated Domain Generalization

The core technical objective of FedDG is to learn feature representations that are stable across source domains and therefore likely to generalize to unseen targets.

• Feature Alignment: Clients collaboratively minimize divergence between their local feature distributions without sharing raw data.
• Distributional Robustness: Models are optimized for worst-case performance across source domains rather than average performance.
• Invariant Risk Minimization (IRM): A common objective that seeks representations where the optimal classifier is identical across all training environments.

To simulate domain shift during local training, each client applies aggressive data augmentation policies that expose the model to stylistic variations.

• Amplitude-Phase Recombination: In medical imaging, clients swap the amplitude spectrum of images to simulate scanner-specific variations.
• Style Transfer Augmentation: Applying random color jitter, blur, and noise to mimic the visual appearance of unseen hospital equipment.
• Distributional Augmentation: Synthesizing adversarial feature statistics during training to force the model to learn domain-agnostic boundaries.

The server orchestrates inter-client knowledge sharing to reduce domain gaps without centralizing data.

• Prototype Sharing: Clients exchange class-conditional feature centroids rather than raw gradients, enabling alignment of decision boundaries across domains.
• Domain Label Disclosure: Each client shares a domain identifier, allowing the aggregation server to explicitly model and compensate for domain-specific biases.
• Cross-Domain Mixup: Virtual training samples are created by interpolating features from different clients, forcing the global model to learn continuous domain transitions.

FedDG systems continuously measure the discrepancy between source and held-out performance to detect overfitting to known domains.

• Leave-One-Domain-Out Validation: Each client is sequentially excluded from training to serve as a proxy for an unseen target domain.
• Domain Divergence Metrics: Tracking Maximum Mean Discrepancy (MMD) or Wasserstein distance between client feature distributions during training.
• Early Stopping on OOD Loss: Training halts when performance on a held-out validation domain begins to degrade, preventing memorization of spurious domain-specific correlations.

FedDG often incorporates gradient-based meta-learning to train a model initialization that can quickly adapt to a new domain with minimal data.

• Model-Agnostic Meta-Learning (MAML): Clients simulate domain shift by splitting local data into meta-train and meta-test sets, optimizing for fast fine-tuning.
• Federated Reptile: A simpler first-order meta-learning algorithm where local updates are treated as steps toward a broadly generalizable initialization.
• Prototypical Networks: Clients learn a metric space where classification is performed by computing distances to class prototypes, naturally generalizing to new classes and domains.

FedDG explicitly models the data heterogeneity that makes standard federated averaging fail under distribution shift.

• Covariate Shift: The input distribution P(X) varies across hospitals due to different scanner vendors and acquisition protocols.
• Concept Shift: The relationship P(Y|X) changes across sites due to varying patient demographics and disease prevalence.
• Label Shift: The prior distribution P(Y) differs across clients, requiring importance-weighted aggregation to avoid biased global models.
• Open Set DG: Some classes present in the target domain may be entirely absent from all source clients, requiring open-set recognition capabilities.

The classical machine learning objective of training a model on multiple source domains that robustly generalizes to an unseen target domain without any prior exposure. Unlike domain adaptation, no target data—labeled or unlabeled—is available during training. The goal is to learn domain-invariant representations that capture the underlying causal structure of the task rather than spurious correlations specific to any single source domain.

A decentralized training paradigm where multiple clients collaboratively learn a shared model without centralizing raw data. Each client trains locally on its private dataset and transmits only model updates—gradients or weights—to a central aggregation server. This architecture preserves data locality and privacy, making it essential for regulated industries like healthcare, but introduces challenges of non-IID data distributions and communication constraints.

The fundamental challenge in federated systems where local client datasets are not independent and identically distributed. In healthcare, this manifests as:

• Label distribution skew: Hospital A treats predominantly cardiac cases while Hospital B specializes in oncology
• Feature distribution skew: Different MRI scanner manufacturers produce varying image characteristics
• Concept drift: The same diagnosis may present differently across demographic populations This heterogeneity directly motivates the need for domain generalization techniques.

A learning paradigm that seeks data representations where the optimal classifier is simultaneously optimal across all training domains. Rather than minimizing average empirical risk, IRM penalizes feature representations that depend on spurious domain-specific correlations. In federated settings, this translates to finding a global model that performs consistently across all client sites without overfitting to any single institution's data biases.

Techniques that explicitly reduce the divergence between feature distributions of different client domains during federated training. Methods include:

• Federated adversarial alignment: Using domain discriminators to encourage domain-invariant features
• Federated moment matching: Aligning first and second-order statistics of feature distributions across clients
• Prototype-based alignment: Sharing class prototypes rather than raw gradients to harmonize representations

The dominant topology for healthcare federated learning where a small number of reliable, stateful clients—typically hospitals or research institutions—participate in training. Unlike cross-device settings with millions of unreliable edge devices, cross-silo assumes:

• Persistent connectivity and full participation in each round
• Substantial local compute resources
• Curated, high-quality datasets with institutional domain characteristics This setting makes domain generalization both more critical and more tractable.