Federated Learning Aggregation
Overview
Federated learning aggregation studies how to combine knowledge from clients that keep their data local, so they can train useful models together. Methods range from FedAvg-style weighted averaging to client weighting, clustering, prototypes, and personalization; newer approaches also align low-rank or function-space updates when model parameters are not directly compatible. Robust filters and trust scores address poisoned or unreliable updates, while secure, asynchronous, and over-the-air aggregation target privacy, delays, and communication costs. No single strategy works best across settings: performance depends on client heterogeneity and participation, and practical designs must trade off accuracy, fairness, security, privacy, and efficiency.
Latest papers
293 papers29 papers in the last four weeks, down 26% on the four weeks before. 0.2% of all new papers.