The School of Computer Science is pleased to present…
Date: Thursday, May 15th, 2025
Time: 10:00 AM
Location: Essex Hall, Room 122
Federated learning enables decentralized model training while preserving user data privacy. However, it suffers from poor convergence and degraded performance under non-independent and identically distributed (non-IID) client data—a common scenario in real-world applications. This thesis proposes SOSFed, a novel aggregation method that leverages a lightweight and architecture-agnostic approximation of the Fisher Information Matrix (FIM) to reweight client updates based on local curvature, termed the Second-Order Signal (SOS). In contrast to existing methods such as FedAvg, which use uniform averaging, and FedFisher, which imposes architectural constraints, SOSFed maintains low computational overhead (O(d)) and broad applicability. We derive theoretical guarantees showing that SOSFed achieves a strictly lower expected global loss than FedAvg as client heterogeneity increases. Experimental results on three benchmark datasets confirm that SOSFed consistently improves test accuracy without compromising privacy. This work contributes a principled and practical approach to enhancing robustness and performance in federated optimization under non-IID settings.
Internal Reader: Dr. Luis Rueda
External Reader: Dr. Abdulkadir Hussein
Advisor: Dr. Dima Alhadidi
Registration Link (only MAC students need to pre-register)