Federated learning lets hospitals train a shared machine learning model without moving patient data. But hospitals see different patients and tests. That mismatch makes a single shared model both hard to adapt to each hospital and brittle when applied to new patient groups. The new paper introduces FedKPer, a way to balance two goals at once: let local models adapt to a hospital’s own data (personalization) while keeping the shared model useful across hospitals (generalization), and to reduce the tendency of models to “forget” patterns learned earlier.

FedKPer works in two linked steps. First, it changes the local training on each site so that a hospital’s model only copies the shared model when that shared model is actually helpful for that hospital’s data. Technically, this is done by adding a “distillation” term that pulls the local model toward the global model, but the strength of that pull is set by how accurate the global model is on the hospital’s own labels. If the global model performs poorly on that site, its influence is weakened. The method also caps that trust level to avoid numerical problems and clips unusually large gradient updates to limit outlier effects.

Second, FedKPer changes how the server combines updates from hospitals into the new shared model. Instead of averaging updates only by dataset size, it gives more weight to local updates that are both reliable (measured by the local training accuracy) and label-diverse (measured by the entropy of the local label histogram). This reduces the risk that hospitals with very skewed class mixes dominate the shared model, which can otherwise push the global model toward biases that hurt general use.

The authors also introduce new ways to measure forgetting. They break the accuracy history into recovery intervals that start at a new peak and end when performance returns to that peak. Those metrics aim to capture whether a method loses previously learned behavior over rounds of training. According to experiments reported in the paper on medical imaging datasets, FedKPer improves the trade-off between personalization and generalization while maintaining retention of past knowledge. The authors provide code and data links for reproducibility.