Structured Models and Algorithms for Sensitive Data
Conor Hassan · Queensland University of Technology · 2024
The use of confidential data in statistical and machine learning models is increasingly prevalent across industries, necessitating the design of privacy-preserving models and algorithms. This thesis addresses this critical challenge through two main approaches: synthetic data generation and federated learning. We develop novel models and algorithms for parameter estimation in scenarios where data is distributed across multiple confidential sources, combining concepts from variational inference, deep generative modeling, and data augmentation. The innovations in this thesis expand the applicability of Bayesian hierarchical models to novel distributed data settings.