Reproduce “Federated Learning by Utility-Constrained Stochastic Aggregation for Improving Rational Participation” on OpenResearch

About this paper

M Yashwanth, Arunabh Singh, Ashok Nayak, Sai Kiran Bulusu, Anirban Chakraborty

FedUCA, a federated learning framework from the Indian Institute of Science and collaborators, sustains client participation by optimizing aggregation weights to meet individual rationality constraints, leveraging a utility-constrained stochastic aggregation method. This approach leads to significantly higher global model accuracy and client participation, for instance, achieving 66.89% accuracy on CIFAR-10 with Dirichlet(0.5, 1.0) heterogeneity, compared to 16.63% for FedAvg.

Key insights

• Modeling client utility as a concave function of its aggregation weight enables the generation of an implicit utility surplus (Jensen's gap) through stochastic aggregation, which acts as a robustness buffer to satisfy participation constraints.
• The Dirichlet concentration parameter (δ) non-monotonically influences client participation by modulating the geometry of the aggregation strategy mixture and, consequently, the induced Jensen surplus.
• Sustaining rational client participation by explicitly addressing individual rationality constraints is crucial for achieving high global model performance and preventing training collapse in heterogeneous cross-silo FL environments.

Summary by alphaXiv.