Client Selection in Federated Learning

Federated learning trains a shared model across devices or institutions that keep their data local, so client selection determines whose updates shape the model in each round. Methods rank or group clients using signals such as gradient or loss, data diversity and uncertainty, device availability and compute cost, or expected contribution; optimization and reinforcement learning can adapt choices over time. Evaluations commonly use image benchmarks such as MNIST and CIFAR alongside edge, IoT, and clinical settings with non-IID data, stragglers, or limited participation. A central tension is that selecting informative and fast clients can speed training but exclude groups or distort the target population, while privacy limits how accurately their data and contributions can be measured.