When Differential Privacy Meets Human Decision-Making
REGISTER
Abstract: This talk presents a series of studies examining how users’ decisions interact with differentially private mechanisms, and how these choices affect the performance of models trained on the data. In an online experiment (n = 734), we show that privacy guarantees significantly shape people’s perceptions of data collection and their willingness to share. We introduce a conjoint-analysis tool that allows data scientists to evaluate and predict these behavioral effects on datasets. A follow-up study (n = 817) reveals that users’ sharing decisions significantly shape the data collected and the performance of a specific machine-learning model. Extending beyond user studies, we also analyze the privacy guarantees (especially epsilon values) used in real-world deployments of differential privacy across academic, commercial, and government settings. Together, these findings reframe the privacy–utility tradeoff: when user behavior and practical implementations are considered, stronger privacy guarantees can paradoxically improve model outcomes.