2022 Theses Master's

PRIVUS: Census Privacy System

Nicita, Alexander

Privus is a Python toolkit for census privacy, instantiated for five countries as well as two privacy libraries, for differential privacy and synthetic data generated via GANs. The Privus toolkit is designed to be both modular and extensible such that other countries' national censuses and other privacy libraries can be seamlessly added to the package. Demonstration experiments comparing the accuracy and privacy trade-offs of various hyper-parameters of each privacy library are included in this paper. Much of the research in this paper has been inspired by the contested decision to apply differential privacy to the 2020 United States census algorithms, driven by the belief that additional exploratory research and software systems development can be a helpful force for understanding the nuances of applying privacy-preserving technologies to national censuses.