Integrating federated learning and differential privacy in dental X-ray for caries segmentation
| dc.contributor.advisor | smith, paul | |
| dc.contributor.author | alharthi, shuruq saad | |
| dc.date.accessioned | 2024-12-29T07:28:26Z | |
| dc.date.issued | 2024-09 | |
| dc.description | This dissertation investigates Federated Learning (FL) as a collaborative AI training technique designed to preserve privacy, with a specific focus on its application in medical imaging, particularly for dental X-ray segmentation in caries detection. Introduced by Google in 2016, FL revolutionized machine learning by enabling model training across multiple devices without the need to transfer raw data, addressing critical privacy concerns in sensitive sectors like healthcare. This study aims to enhance the security and effectiveness of FL by incorporating differential privacy techniques to protect patient data while preserving model accuracy. The research systematically evaluates FL’s performance with varying numbers of clients, examines the impact of unbalanced data distributions, and applies differential privacy using different noise levels to improve data security. The experiments demonstrate that FL can effectively safeguard sensitive information, offering a decentralized alternative to traditional centralized machine learning methods, which carry significant risks due to data centralization. While FL addresses key privacy concerns, the study acknowledges that security issues remain, and the development of sophisticated safeguards against potential attacks is outside the scope of this research. The findings contribute to the understanding of FL’s potential in healthcare, highlighting its ability to balance privacy preservation with collaborative AI training. This research underscores FL’s viability as a privacy-preserving technology, providing valuable insights into optimizing AI models in medical applications without compromising data security. A key outcome of the experiment is that FL with a higher number of clients performs more effectively than models with fewer clients. Additionally, random data distribution outperformed other methods in most of the evaluated metrics. | |
| dc.description.abstract | . | |
| dc.format.extent | 85 | |
| dc.identifier.uri | https://hdl.handle.net/20.500.14154/74504 | |
| dc.language.iso | en | |
| dc.publisher | Lancaster university | |
| dc.subject | Decentralized Machine Learning | |
| dc.subject | Privacy-preserving AI | |
| dc.subject | Federated Learning (FL) | |
| dc.subject | Differential Privacy | |
| dc.subject | Medical Imaging | |
| dc.title | Integrating federated learning and differential privacy in dental X-ray for caries segmentation | |
| dc.type | Thesis | |
| sdl.degree.department | Communication and computation | |
| sdl.degree.discipline | Cyber security | |
| sdl.degree.grantor | Lancaster university | |
| sdl.degree.name | master |