Eureka: A General Framework for Black-box Differential Privacy Estimators
Yun Lu, Malik Magdon-Ismail, Yu Wei, Vassilis Zikas
IEEE Symposium on Security and Privacy 2024 · Day 1 · Continental Ballroom 6
In the realm of data privacy, ensuring that algorithms do not inadvertently leak sensitive information is paramount. This talk introduces "Eureka," a novel and general framework designed for **black-box differential privacy (DP) estimators**. Presented by Yun Lu, a PhD student at Purdue University, alongside collaborators Malik Magdon-Ismail, Yu Wei, and his advisor Vassilis Zikas, the work addresses a critical gap: enabling domain experts without specialized privacy knowledge to empirically assess the privacy guarantees of their own machine learning mechanisms.
AI review
This talk introduces Eureka, a novel black-box framework for tightly estimating differential privacy guarantees, a critical capability missing from existing DP auditing tools. Its core innovation lies in reducing the complex privacy quantification problem to a solvable binary classification task, enabling non-experts to accurately assess their mechanisms. This is a clever and highly impactful piece of research that significantly lowers the barrier for robust DP adoption.