The School of Computer Science is pleased to present…
Robust, Poisoning-Resilient Federated Learning
MSc Thesis Defense by: Sayedali Sheykholeslamzadeh
Date: Friday, September 5th, 2025
Time: 10:00 AM
Location: Essex Hall, Room 122
Federated Learning (FL) enables collaborative model training without centralizing data, offering strong privacy benefits. However, its practical deployment is still challenged by two critical issues. First, statistical heterogeneity, where clients possess data with highly skewed distributions, can slow or destabilize training. Second, malicious participants may inject poisoned updates, compromising model integrity without requiring detectable anomalies or auxiliary verification data. This thesis addresses both challenges through two complementary methods. The first, SOSFed, introduces a curvature-aware aggregation strategy that gives greater influence to informative client updates, improving convergence and accuracy in heterogeneous settings with minimal computational overhead. The second, PACT, is a passive and validation-free detection approach that evaluates the trustworthiness of clients based on the impact of their updates on model performance, enabling effective identification of malicious participants even when they form the majority. Extensive experiments on benchmark datasets demonstrate that SOSFed consistently improves accuracy over established baselines, while PACT achieves the highest detection rate under strong targeted attacks, all with modest runtime cost. Together, these methods provide a practical toolkit for enhancing both robustness to data skew and resilience against poisoning, moving FL closer to a reliable, real-world deployment.
Internal Reader: Dr. Luis Rueda
External Reader: Dr. Abdul A. Hussein
Advisor: Dr. Dima Alhadidi
Chair: Dr. Muhammad Asaduzzaman
