Reproduce “Federated Learning with Convex Global and Local Constraints” on OpenResearch

About this paper

Chuan He, Le Peng, Ju Sun

Researchers from the University of Minnesota developed the first federated learning framework capable of handling general convex global and local constraints, achieving an $O(\max\{\epsilon_1^{-2}, \epsilon_2^{-2}\})$ outer iteration complexity. This framework empirically matches centralized methods on tasks like Neyman-Pearson classification and fairness-aware learning, demonstrating its practical efficacy.

Key insights

• General convex global and local constraints can be effectively addressed within a federated learning framework by adapting the proximal Augmented Lagrangian method and using an ADMM-based sub-solver.
• Theoretical worst-case complexity guarantees for constrained federated learning are achievable, even under the more practical assumption of locally Lipschitz continuous gradients.
• A well-designed inexact ADMM sub-solver with a novel, communication-light termination criterion can ensure global linear convergence while minimizing communication overhead in the federated setting.

Summary by alphaXiv.