CA-MMDS framework enables asynchronous federated learning with evolving clients and labels

Traditional federated learning assumes fixed clients, data, and objectives, but real-world federations evolve over time—new clients join, existing clients change, and label spaces expand. This breaks conventional round-wise model aggregation, which requires repeated communication, local computation, and synchronized participation from all accumulated clients. Can Peng and colleagues address this with CA-MMDS, a continual multiple-model distillation framework. Instead of aggregating parameters from all clients every round, CA-MMDS keeps a server-side archive of local models and updates the global model through proxy-based distillation across archived models. When new clients appear or existing ones update, only the changed models are uploaded; unchanged clients stay offline yet continue contributing via their archived models.

The framework substantially reduces communication and computation costs while enabling flexible asynchronous cooperation among evolving clients. In experiments on multi-class 3D abdominal CT segmentation, CA-MMDS efficiently incorporated evolving client knowledge and achieved competitive segmentation performance. The work tackles a practical bottleneck in federated learning—handling non-stationary client sets and label spaces without forcing all participants into synchronous rounds. This makes CA-MMDS particularly relevant for healthcare AI, where hospitals (clients) frequently update their data, add new diagnostic categories, or join/leave collaborations. The method retains the privacy benefits of federated learning while eliminating the overhead of constant re-aggregation.