An Interactive Framework for Finding the Optimal Trade-off in Differential Privacy

Abstract: Differential privacy (DP) is the gold standard for privacy-preserving analysis but introduces a fundamental trade-off between privacy guarantees and model performance. Selecting the optimal balance is a critical challenge, framed as a multi-objective optimization (MOO) problem of discovering the Pareto front and eliciting a decision-maker's preference. While interactive MOO offers a solution, standard approaches---which model objectives separately and rely on simple pairwise feedback---are suboptimal for DP because they do not utilize problem structure. In this work, we propose a method, \textbf{PACE} (\textbf{P}rivacy-\textbf{A}ccuracy \textbf{C}urve \textbf{E}licitation), that exploits two key properties to reduce this inefficiency. First, we leverage the fact that the privacy level naturally serves as a constraint: maximizing accuracy for a fixed privacy level generates a solution on the Pareto front. Second, to efficiently model this trade-off, we theoretically derive the trade-off shape for regularized logistic regression, revealing a characteristic S-curve. This theoretical grounding motivates us to model the Pareto front using a sigmoidal function. We empirically demonstrate its effectiveness across studied DP settings. This model allows us to replace less efficient pairwise comparisons with a richer interaction scheme where decision-makers directly select their most preferred solution from the hypothetical trade-off curve. Experiments on differentially private logistic regression and deep transfer learning across six datasets show that PACE converges to the most preferred trade-off with fewer model evaluations and interactions than baselines.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Elliot_Creager1

Submission Number: 8201