FedPer

FedPer divides the neural network architecture into two distinct segments: base layers (feature extractors) and personalization layers (classification heads). The base layers, typically comprising convolutional or early dense layers, learn generalizable representations from the aggregated knowledge of all clients. The final classification layers remain local and are never shared with the central server. This structural separation allows each clinical site to maintain a site-specific decision boundary while benefiting from a globally robust feature space.

Unlike statistical heterogeneity (differences in data distribution), structural heterogeneity refers to differences in the feature space or label space across clients. In healthcare, this manifests as:

• Varying label granularity: Hospital A classifies diseases at a coarse level, while Hospital B uses fine-grained sub-types.
• Different diagnostic criteria: Institutions may use distinct clinical coding standards. FedPer resolves this by allowing each client to maintain a personalized classification layer tailored to its specific label ontology, while the shared base layers learn universal medical imaging features.

By transmitting only the base layer parameters during federated aggregation rounds, FedPer significantly reduces communication overhead. The personalized classification heads, which can be substantial in networks with many output classes, never leave the local client. This design is particularly advantageous in bandwidth-constrained hospital environments where large model uploads may disrupt critical clinical network operations. The reduction in transmitted parameters directly correlates with faster round times and lower infrastructure costs.

In standard federated learning, aggressive global aggregation can overwrite locally specialized knowledge, a phenomenon known as catastrophic forgetting. FedPer's architecture inherently protects local expertise by isolating the personalized layers from the aggregation process. The base layers converge on a universal feature representation that generalizes across sites, while the frozen local heads preserve site-specific diagnostic nuances. This dual structure ensures that a model trained to detect rare conditions at a specialized clinic does not lose that capability after global synchronization.

The FedPer training loop operates in distinct phases:

• Local Phase: Each client performs full forward and backward passes, updating both base and personalization layers using its private data.
• Communication Phase: Only the updated base layer weights are transmitted to the central server.
• Aggregation Phase: The server applies a standard aggregation algorithm, such as Federated Averaging (FedAvg), to the received base layers.
• Distribution Phase: The aggregated base layers are broadcast back to all clients, where they are recombined with the unchanged local personalization heads for the next round.

Unlike Local Fine-Tuning, which adapts the entire model post-federation and risks overfitting to small local datasets, FedPer constrains personalization to the final layers. This provides a regularization effect that prevents the model from diverging too far from the globally learned features. Compared to Multi-Task Federated Learning, FedPer does not require explicit task relationship modeling. The shared base layers implicitly learn a task-agnostic representation, making the architecture simpler to implement and tune in production clinical environments.

A closely related algorithm that also partitions the neural network into shared base layers and personalized head layers. Unlike FedPer, which only personalizes the classifier, FedRep alternates training between learning a global feature extractor and local heads over multiple local epochs, often leading to better representation quality in highly heterogeneous vision tasks.

A generalization of the FedPer concept that systematically determines which layers to personalize and which to keep global. Strategies include:

• Personalizing batch normalization layers to handle domain shift
• Selecting layers based on Fisher information sensitivity
• Dynamic layer selection based on local data size This approach reduces communication overhead compared to full-model personalization while maintaining accuracy.

The post-aggregation process where a globally trained model is further adapted on a specific client's local dataset. FedPer formalizes this by architecturally designating the final classification layers as local, ensuring that fine-tuning occurs only in a controlled subset of parameters. This prevents the catastrophic forgetting of global features while allowing site-specific decision boundaries.

FedPer directly addresses label distribution skew—a core non-IID challenge where different hospitals have varying proportions of disease classes. By keeping base layers synchronized to learn universal medical feature extractors and allowing local heads to calibrate to site-specific class priors, FedPer mitigates the weight divergence that plagues standard FedAvg in heterogeneous clinical environments.

FTL addresses scenarios where clients have different feature spaces or different label spaces. FedPer complements FTL by providing a structural template: the shared base layers can be transferred across institutions with overlapping feature modalities (e.g., chest X-rays), while the personalized head accommodates institution-specific diagnostic label sets without requiring a unified global taxonomy.

A personalization technique that mixes the parameters of a local model and a global model to find an optimal balance point. FedPer can be viewed as a hard structural interpolation where the interpolation weight is 1.0 for the head layers and 0.0 for the base layers. More flexible interpolation methods like APFL learn a continuous mixing coefficient per layer, offering a spectrum between FedAvg and FedPer extremes.