Balancing privacy, accuracy, and communication in federated learning

Federated learning trains models across distributed devices, but existing methods sacrifice either accuracy (FedAvg's bias) or communication efficiency (FedSGD's cost). This work proposes FedHybrid, which combines FedAvg initialization with FedSGD iterations, and FedNewton, which averages local Newton steps to reduce bias. The authors derive finite-sample error bounds for differentially private versions of these estimators and establish a minimax lower bound that benchmarks optimality. Experiments on logistic regression and neural networks on MNIST and CIFAR-10 validate the approach.