Privacy-Preserving Federated Analysis for Decentralized Health Data Analysis

Sammanfattning

This thesis examines the balance between protecting individual privacy and accuracy while gaining insights from three health datasets and one finance dataset using differential privacy (DP).The study analyzes these datasets through statistical and machine learning methods, while strictly maintaining privacy requirements. A key focus of this research is differential privacy, which provides a mathematical guarantee that individual privacy is preserved during statistical analysis. The study investigates how varying the privacy level, represented by the parameter epsilon, affects the accuracy of important statistical estimates like the mean and variance. A key contribution of this research is the analysis of health and finance data distributed across multiple edge nodes—local servers that process data closer to where it is generated. This study evaluates how the privacy-utility trade-off shifts when each node’s data is kept private. The research further explores secure multiparty computation (SMC), which enables parties to compute functions together without revealing individual data. This approach requires only global differential privacy, ensuring that no single participant can see others’ data. Numerical experiments demonstrate that adequate accuracy can be achieved with acceptable privacy using secure aggregation. For instance, without secure aggregation, a federation of 8 participants needs to allow for a 10 times larger privacy risk parameter, while a federation of 64 participants requires nearly 40 times larger. The conclusion is that combining secure aggregation with global differential privacy offers the best balance between privacy and data utility across all datasets.