| Federated Learning (FL) enables collaborative model training across distributed clients without transferring raw data, thereby meeting stringent privacy standards. Non-practical FL systems are, however, characterized by high communication costs, unfair participation of clients, as well as suboptimal convergence with non-IID data distributions. Random selection of clients does not offer any explicit guarantees of fairness, whereas heuristic approaches can create bias of participation. We present FedBIBD, a federated learning model that leverages Balanced Incomplete Block Design (BIBD) theory to overcome these difficulties. To promote balanced participation, control client co-occurrence to minimize gradient variance, and limit communication costs by setting a fixed number of clients per round, FedBIBD uses deterministic client scheduling. We achieve a convergence rate of O(1/√T ) for non-convex optimization problems and provide an efficient stochastic algorithm with fixed client memory and linear server aggregation. Experiments with MNIST, Fashion-MNIST, and CIFAR-10 show that FedBIBD is a significant advancement in fairness and communication overheads, achieving competitive or even superior performance to other federated learning algorithms. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.