Federated Normalization Aggregation (FedBN)

The core innovation of FedBN is the exclusion of batch normalization (BN) layers from the global aggregation process. While convolutional and fully connected layers are averaged across clients to learn shared feature representations, BN layers remain locally private and are never shared with the server. This allows each participating institution to maintain its own running mean and variance statistics that capture site-specific data characteristics, such as differences in scanner vendors, imaging protocols, or patient demographics. The result is a model that learns globally shared feature extractors while preserving locally adaptive normalization.

FedBN is specifically designed to address feature shift, a form of statistical heterogeneity where the marginal distribution of input features P(x) differs across clients, even if the conditional label distribution P(y|x) remains similar. In medical imaging, this manifests as:

• Scanner variability: Different MRI machines produce images with varying intensity distributions
• Protocol differences: Institutions use different acquisition parameters or contrast agents
• Population demographics: Patient age, ethnicity, or comorbidity profiles shift the input distribution By keeping BN statistics local, FedBN prevents the global model from averaging away these meaningful domain-specific variations, allowing each site's model to normalize inputs according to its own data distribution.

Empirical studies demonstrate that FedBN achieves significantly faster convergence compared to standard FedAvg when training across domains with feature shift. The mechanism behind this acceleration is straightforward: when BN statistics are forcibly averaged across clients with different data distributions, the global model receives conflicting normalization signals that create optimization turbulence. By isolating BN layers, FedBN eliminates this source of gradient interference, allowing the shared layers to focus on learning domain-invariant features while local BN layers handle domain-specific scaling. This decoupling reduces the number of communication rounds required to reach target accuracy by 30-50% in typical cross-silo medical imaging scenarios.

FedBN provides an implicit form of unsupervised domain adaptation without requiring any data sharing between institutions. Since BN statistics encode the first and second moments of each client's data distribution, keeping them local means that no distributional information about a site's patient population is ever transmitted to the server or other participants. This is particularly valuable in healthcare settings where:

• The statistical properties of a dataset can reveal demographic information about the patient cohort
• Institutions may consider their data distribution characteristics as proprietary or sensitive
• Regulatory frameworks like GDPR may classify distributional metadata as protected information FedBN achieves domain adaptation as a natural byproduct of its architecture rather than through explicit privacy mechanisms.

FedBN requires minimal modification to existing federated learning pipelines and is compatible with most modern neural network architectures that use batch normalization. The implementation involves:

• Identifying BN layers in the model architecture before training begins
• Partitioning model parameters into shared parameters (weights and biases of conv/linear layers) and local parameters (BN running means, variances, scales, and shifts)
• Modifying the aggregation step to only average shared parameters while skipping BN parameters This approach is orthogonal to other federated optimization improvements such as FedProx, SCAFFOLD, or differential privacy mechanisms, meaning FedBN can be combined with these techniques for compounded benefits without architectural conflicts.

FedBN has shown particular efficacy in cross-silo healthcare deployments where multiple hospitals collaborate on computer vision tasks. In benchmark evaluations using multi-site medical imaging datasets:

• Brain MRI segmentation across different scanner vendors showed 8-12% improvement in Dice coefficient compared to FedAvg
• Chest X-ray classification across hospitals with different imaging equipment demonstrated reduced performance variance across sites
• Histopathology analysis with different staining protocols benefited from local BN statistics capturing color distribution differences The technique is especially valuable when the number of participating institutions is small (3-10), which is typical in healthcare consortia, as each site's domain characteristics have outsized influence on the global model.