Context-Aware Client Selection in Centralized Federated Learning
PhD Dissertation Defense by: Soroush Ziaeinejad
Date: September 10, 2026
Time: 10:00 AM
Location: Essex Hall 122
Abstract:
Federated Learning (FL) is a decentralized machine learning paradigm that enables multiple data holders, referred to as clients, to collaboratively train a shared model without transferring their raw data to a central server. A central design decision in FL is client selection, the process of determining which subset of clients participates in each training round. Most existing FL systems rely on random sampling or single-metric selection strategies, which do not account for the contextual characteristics of clients and can lead to slow convergence, unstable training, and biased global models. This dissertation studies context-aware client selection in centralized FL, where the server incorporates contextual information about clients and their environments, such as data characteristics, system capabilities, and participation history, into the selection process. It surveys the field, proposes a set of context-aware client collaboration strategies, and applies them to heterogeneous medical imaging and to the federated fine-tuning of Large Language Models. The findings demonstrate that lightweight and privacy-preserving contextual signals can meaningfully improve the performance, stability, and fairness of FL systems without altering model architectures or aggregation algorithms.
Keywords:
Federated Learning, Privacy-preserving AI, Distributed Learning, Client Selection, Data Heterogeneity
Doctoral Committee:
Internal Reader: Dr. Jianguo Lu
Internal Reader: Dr. Muhammad Asaduzzaman
External Reader: Dr. Majid Ahmadi
External Examiner: Dr. Ali Miri
Advisor: Dr. Saeed Samet
Chair: Dr. Michael Stasko