Federated Prototype Learning

Instead of averaging model weights, FPL aggregates class prototypes—compact vector representations of each class learned from local data. Clients compute the mean embedding for each class and transmit only these statistical summaries to the server. The server averages prototypes from all clients to form global prototypes, which capture the central tendency of each class across the entire federated network. This approach dramatically reduces communication overhead while preserving privacy, as raw data never leaves the local site.

Global prototypes are redistributed to clients and used as regularization targets during local training. The local loss function is augmented with a proximity term that penalizes the distance between local class embeddings and their corresponding global prototypes. This constraint prevents local models from drifting too far toward their own biased data distributions, effectively correcting label distribution skew where certain clients have highly imbalanced class representations. The regularization strength can be tuned per-client to balance personalization with global consistency.

Unlike weight-averaging methods like FedAvg, FPL naturally supports heterogeneous model architectures across clients. Since only prototypes—fixed-dimensional embedding vectors—are exchanged, clients can use different neural network backbones suited to their computational resources. A rural clinic running a lightweight MobileNet can collaborate with an academic medical center using a ResNet-152, as long as both produce embeddings in the same shared representation space. This architectural agnosticism makes FPL ideal for real-world healthcare deployments with diverse hardware capabilities.

FPL explicitly addresses the common federated challenge where clients possess different label distributions. When a client lacks examples of a rare disease class, the global prototype for that class still arrives from other clients that have it. The local model is trained to minimize distance to this absent-class prototype, enabling it to learn discriminative features for classes it has never seen locally. This mechanism provides a form of zero-shot knowledge transfer across the network, dramatically improving performance on rare conditions in data-poor sites.

FPL achieves significant bandwidth reduction compared to gradient-based methods. A typical ResNet-50 has over 25 million parameters requiring transmission per round in FedAvg. In contrast, FPL transmits only C × D floating-point values per client, where C is the number of classes and D is the embedding dimension (typically 128-512). For a 10-class problem with 256-dimensional prototypes, this amounts to just 2,560 values—a reduction of roughly 10,000× in communication payload. This efficiency enables participation from bandwidth-constrained edge devices and reduces carbon footprint.

FPL provides inherent privacy advantages by design. Clients never share raw data, model weights, or gradients—only aggregated class-level statistics. The prototypes represent averages over all samples of a class, making individual patient reconstruction infeasible. For enhanced protection, differential privacy can be applied by adding calibrated noise to prototypes before transmission. Unlike gradient-based methods vulnerable to deep leakage attacks, prototype inversion attacks yield only blurry class averages rather than specific training examples, providing a stronger privacy-utility tradeoff for healthcare applications.