Tools
Tools
Yeddla, Naga Sai Bhaskar Naveen (2024) Federated Learning and Privacy-Preserving Artificial Intelligence. Masters thesis, Dublin, National College of Ireland.
Preview |
PDF (Master of Science)
Download (1MB) | Preview |
Preview |
PDF (Configuration Manual)
Download (564kB) | Preview |
Abstract
This research integrates federated learning with privacy-preserving techniques, specifically differential privacy and homomorphic encryption, to enhance credit card fraud detection. Traditional models are generally centralized and, therefore, suffer from considerable challenges related to privacy risk and regulatory compliance, including GDPR. Federated learning is a decentralized approach whereby models can be trained across distributed datasets without sharing raw data. This paper analyzes two types of financial transactional datasets one real and one artificial using machine learning approaches, including random forests and gradient boosting. The study examines how neural network methods are applied in both federated and centralized data settings. The key findings present strong fraud detection rates using Federated Supervised Deep Learning (FSDL), almost identically for all datasets. This approach also provides improved data confidentiality and privacy security. Enhanced methods include differential and homomorphic encryption; these provide robust protection but with higher computational costs. These findings point out the urgent need for optimization to reduce computational overhead. Therefore, this work is a very good trade-off between fraud detection performance and compliance with the constraints of privacy. Thus, this will be the contribution of this work towards ethical AI applications in finance.
Actions (login required)
 |
View Item |