FedProx

FedProx is a robust federated learning framework designed to handle systems and statistical heterogeneity through a proximal term that stabilizes local updates.

FedProx solves the core instability issues in decentralized machine learning by introducing a proximal term to the standard FedAvg objective. This mathematical constraint limits the impact of local updates, preventing divergent global models when device hardware (systems heterogeneity) or data distributions (statistical heterogeneity) vary wildly. Developed by researchers at Carnegie Mellon University, the framework allows participants to perform variable amounts of work based on their available compute power. In large-scale benchmarks like the LEAF datasets, FedProx demonstrates superior convergence stability over FedAvg, particularly in networks with high percentages of stragglers or non-IID data.

https://arxiv.org/abs/1812.06127