Evaluating the Transferability of Data Poisoning Attacks across Centralized & Federated Learning Techniques

This research investigated the main differences between Centralized Learning (CL) and Federated Learning (FL) in the context of data poisoning attacks focusing on how each architecture mitigated or stopped adversarial disruptions. Computer vision and classification tasks within machine learning models across diverse datasets was used. Evaluation around the performance of CL and FL under different levels of data poisoning was done to provide a comprehensive understanding of how learning technique affected overall vulnerability. Across the trials, special focus was paid to evaluating how model accuracy, precision, recall, cross entropy loss, and F1 score changed as the level of data poisoning increased with different datasets. Different models and datasets were tested under controlled poisoning situations to simulate real-world environments building a deeper understanding of the tension of data security and design in modern machine learning (ML) systems, largely classification tasks within computer vision.The importance of this is shown by the growing adoption of FL as a way of enabling Distributed Artificial Intelligence (DAI), especially as organizations aim to expand AI integration while making data privacy stronger against adversarial threats. Key operational considerations driving this shift include managing data with differing classification or control levels, enforcing role-based data access, and maintaining performance consistency despite the increasing geographic dispersion of data sources (an inherent limitation of centralized architectures that we see today). In exploring both the strengths and weaknesses of each learning technique, this research informs more secure and context-appropriate AI system design across varying use cases and industries.  

Author

• Mattarock II, Michael J.

Language

• en

Keyword

• ML Operations
• Federated Learning
• Distributed AI
• Data Poisoning
• Adversarial Training
• AI

Date created

• 2025

Type of Work

• Dissertation

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Cybersecurity Analytics

Degree

• D.Engr.

Advisor

• Dargan, John

Committee Member(s)

• Clements, Ken
• Darko, Michael

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/br86b4419