Towards Trustworthy Federated Learning

Author

Alina Basharat, The University of Texas Rio Grande Valley

Date of Award

12-1-2025

Document Type

Thesis

Degree Name

Master of Science in Engineering (MSE)

Department

Electrical Engineering

First Advisor

Rogelio Soto

Second Advisor

Ping Xu

Third Advisor

Wenjie Dong

Abstract

Federated learning is a collaborative training model in which multiple clients optimize aglobal model by transmitting updates to a coordinating server while keeping raw data on-device, thereby reducing direct data exposure and enabling iterative global improvement. However, the iterative communication process is vulnerable to malicious attackers that either deliberately destroy the model or curious to infer raw data. Moreover, learning from multiple agents may result in unfair results. To enhance trustworthiness within this setting, we employ two-sided norm-based screening (TNBS) that removes both abnormally large and abnormally small updates, pair it with a q-fair objective to emphasize high-loss (disadvantaged) clients, and enforce client-level differential privacy with (ε, δ ) guarantees. For hierarchical deployments, TS-FedNBS introduces a two-stage defense for federated edge computing: norm based screening at edge nodes and robust median aggregation at the server. Experiments on benchmark datasets show resilience to Gaussian, inner-product manipulation, and omniscient attacks, approaching no-attack accuracy and highlighting the necessity of server-side robustness in hierarchical systems. Finally, Fed-ETF advances Byzantine resilience via cosine-based trust scoring relative to a server reference, selective participation with trust-weighted aggregation, and EMA-based adaptation of client learning rates. Different experiments guarantees under non-IID data with adversarial presence are performed to show the performance of the method.

Comments

Copyright 2025 Alina Basharat. All Rights Reserved. https://proquest.com/docview/3292599847

Recommended Citation

Basharat, A. (2025). Towards Trustworthy Federated Learning [Master's thesis, The University of Texas Rio Grande Valley]. ScholarWorks @ UTRGV. https://scholarworks.utrgv.edu/etd/1829