Dissertation Title:  Advanced Federated Learning Algorithms Leveraging Selective Forgetting for Data-Constrained Environments

Abstract:
Federated Learning (FL) has emerged as a pivotal approach for machine learning, enabling model training across clients without requiring data to be centrally aggregated. However, in real-world applications, insufficient and non-IID training data can lead to suboptimal model performance. Knowledge Evolution (KE) and Later-Layer Forgetting (LLF), inspired by the Lottery Ticket Hypothesis (LTH), have demonstrated significant accuracy gains in limited data settings by iteratively pruning and evolving neural networks to identify high-performing subnetworks. While KE and LLF offer compelling solutions for improving model training, they are typically designed for centralized data scenarios, which limits their direct applicability in decentralized and privacy-preserving environments. 

In this dissertation, we propose novel frameworks—Federated Learning with Knowledge Evolution (FL-KE) and Federated Learning with Later-Layer Forgetting (FL-LLF)—to address the challenge of insufficient training data in federated settings. These frameworks aim to enhance model generalization and convergence by evolving subnetworks over generations and strategically resetting deeper layers of neural networks. To evaluate the security and robustness of these frameworks, we analyze their resilience against classical FL attack vectors, including poisoning, backdoor, and inference attacks, demonstrating their potential as defense mechanisms in adversarial settings. 

Furthermore, we complement our technical contributions with a regulatory perspective by conducting a survey targeting key stakeholders in the Gulf Cooperation Council (GCC) region. The survey investigates how organizations balance compliance with emerging data protection laws and regulations while mitigating real-world threats against federated learning systems. By bridging methodological innovation with empirical insights on regulatory alignment, this thesis contributes a holistic framework for secure, scalable, and privacy-compliant federated machine learning.

Committee members:

Lalana Kagal
Principal Research Scientist, MIT Computer Science and Artificial Intelligence Lab

Sharifa Alghowinem
Research Scientist, MIT Media Lab 

Praneeth Vepakomma
Assistant Professor, Mohamed bin Zayed University of Artificial Intelligence
Visiting Assistant Professor, MIT Institute for Data, Systems, and Society