Federated Prototype Learning

Instead of transmitting high-dimensional model weights or gradients, each client computes and shares class prototypes—the mean embedding vector for each category in the latent space. This reduces communication payloads by orders of magnitude compared to standard Federated Averaging, as only a few vectors per class are transmitted rather than millions of parameters.

Because prototypes exist in a shared latent representation space rather than parameter space, participating clients can use entirely different model architectures locally. A hospital with a ResNet-50 can collaborate with another using a Vision Transformer, as long as both project inputs into the same-dimensional embedding space. This architectural agnosticism is critical for real-world healthcare consortia with diverse infrastructure.

Sharing abstract prototypes rather than raw data or gradients provides a natural privacy barrier. Prototypes are aggregated statistical summaries of local class distributions, making direct data reconstruction significantly harder than with gradient-based methods. When combined with differential privacy, prototype noise injection can provide formal privacy guarantees without destroying semantic utility.

The server collects local prototypes from all clients and computes global prototypes via weighted averaging, typically proportional to each client's sample count per class. These global prototypes are then redistributed, and local models are regularized to align their embeddings with the global representatives. This creates a shared semantic understanding without ever centralizing patient data.

Prototype learning demonstrates strong resilience to statistical heterogeneity across clinical sites. Even when label distributions are skewed—one hospital has many positive cases while another has few—the global prototypes converge to representative class centroids. Local models use a prototype-based loss that minimizes distance to the correct class prototype while maximizing separation from others, naturally handling class imbalance.

Local training incorporates a prototype-based regularization term alongside standard supervised loss. This term penalizes the distance between a sample's embedding and its corresponding global prototype, while encouraging separation from prototypes of other classes. This dual objective prevents local models from diverging during isolated training rounds and accelerates global convergence.

A classification paradigm where each class is represented by a prototype vector—a learned embedding that captures the central tendency of that category. During inference, a query sample is assigned to the class whose prototype is nearest in the embedding space.

• Prototypes are computed by averaging the embeddings of all support samples belonging to a class
• Distance metrics like Euclidean or cosine similarity determine class membership
• In federated settings, clients share only these abstract prototypes rather than raw data or full model gradients
• This approach provides inherent interpretability, as prototypes can be visualized or inspected

The core aggregation mechanism in federated prototype learning where local class prototypes computed on each client are securely averaged on a central server to form global prototypes.

• Each client computes the mean embedding vector for each class present in its local data
• Prototypes are transmitted to the server instead of model weights or gradients
• The server performs weighted averaging based on the number of samples each client used per class
• Global prototypes are redistributed to all clients for the next round of local adaptation
• This dramatically reduces communication overhead compared to FedAvg, as prototypes are compact vectors

A neural architecture consisting of an embedding function that maps input samples into a metric space where classification is performed by computing distances to class prototypes.

• The embedding function is typically a convolutional neural network for images or a transformer for sequential data
• The network is trained episodically on few-shot learning tasks to learn a generalizable metric space
• In federated prototype learning, the embedding network is trained locally while prototypes are shared globally
• The architecture supports open-set recognition, where new classes can be added by simply computing their prototypes without retraining

Sharing class prototypes provides stronger privacy protection than gradient sharing because prototypes are aggregated statistics that obscure individual sample contributions.

• Prototypes are mean vectors computed over multiple samples, making membership inference attacks significantly harder
• Differential privacy can be applied by adding calibrated Gaussian noise to prototypes before transmission
• The information bottleneck of compressing a class into a single vector naturally limits data leakage
• Prototypes are less susceptible to gradient inversion attacks that can reconstruct training data from shared model updates
• Formal privacy accounting tracks the cumulative privacy loss across federated rounds

A key challenge in federated prototype learning where clients have non-overlapping class sets or highly imbalanced class distributions, requiring specialized aggregation strategies.

• Clients may only possess data for a subset of the global classes, creating partial prototype contributions
• The server must handle missing prototypes from clients that lack certain classes
• Class-incremental learning techniques allow the global model to incorporate new classes discovered at individual sites
• Prototype normalization and alignment procedures correct for distribution shift between client populations
• Weighted aggregation accounts for the statistical reliability of prototypes based on local sample counts