핵심 개념
Conformal prediction framework for privacy and uncertainty quantification.
초록
The content introduces a framework for private prediction sets using conformal prediction to address privacy and uncertainty quantification in machine learning systems. It discusses the challenges of combining privacy and uncertainty, presents algorithms, theoretical guarantees, and empirical evaluations. Key highlights include:
Introduction to the need for privacy and uncertainty quantification in machine learning systems.
Framework based on conformal prediction for private prediction sets.
Methodology for differential privacy in prediction sets.
Theoretical guarantees on coverage and privacy.
Empirical evaluation on image classification problems and COVID-19 diagnosis.
통계
"ϵ-differentially private (1 − α + O((nϵ)−1))-quantile of {si}n i=1, denoted ˆs"
"ϵ-differentially private algorithm A"
"ϵ-differentially private mechanism for fitting C"
인용구
"We present a framework that treats these two desiderata jointly."
"Our main contribution is a privacy-preserving algorithm."