Angela Mitrovska: Systematic Evaluation of Different Learning Scenarios for MRI-based Alzheimer’s Disease Detection using Federated Learning with Secure Aggregation

BCCN Berlin / Technische Universität Berlin

Abstract

The rapid advancement of machine learning has also been reflected through the development of AI-based methods in the field of neuroimaging. However, the development of novel AI-based methods in the field, is set back by the lack of publicly-available large scale datasets, due to the ethical and legal requirements to protect the patient’s privacy. The publicly-available datasets are usually very small, and stem from very few institutions and geographic regions, possibly leading to unquantifiable bias with respect to co-variables. Privacy-preserving AI offers techniques to help bridge the gap between the protection of the patient’s data and data utilization. Federated Learning, a method under privacy-preserving AI, allows for collaborative data utilization, without having direct access to the patient’s data. The thesis demonstrates the training of a machine learning model for Alzheimer’s Disease detection in a federated setting. As Federated Learning, is not fully privacy-preserving by itself, its implementation is extended with the addition of Secure Aggregation, a privacy-preserving method based on Secure Multi-Party Computation. The thesis explores how models trained using Federated Learning with, and without Secure Aggregation, compare to models trained in the conventional setting (i.e. centralized approach) in different heterogeneous environments where different clients posses data with different statistical and demographic distributions. Particular attention is paid to the case where data is non-IID across the clients participating in Federated Learning. These differences are investigated, to explore their effect on the performance of the models trained in a federated setting. Additionally, the thesis simulates membership and attribute inference attacks, to quantify the information leakage and explore the privacy guarantees of models trained using Federated Learning with, and without Secure Aggregation.

 

Additional Information

Master Thesis Defense

 

Organized by

Prof. Dr. Kerstin Ritter   & Prof. Dr.-Ing. Ronald Freund     / Lisa Velenosi