Estimating and Auditing (f-)DP
- 講者呂昀Yun Lu 教授 (加拿大維多利亞大學)
邀請人：鐘楷閔 - 時間2025-01-24 (Fri.) 10:15 ~ 12:15
- 地點資訊所新館101演講廳
摘要
Differential privacy (DP) is currently the most widely-used measure of privacy. However, there are several challenges in the deployment of DP. First, the DP privacy definition has well-acknowledged limitations, such as difficulty in describing the privacy of multiple composed mechanisms. This prompted several generalisations of DP, one of the most recent and well-discussed of which is f-DP, which has several attractive properties including `lossless’ composition. The second difficulty in practical deployments is that DP is hard to verify, especially in the black-box setting. This has led to a plethora of works that aim to (empirically) either directly `estimate’ the level of privacy of an implementation, or `audit’ whether it actually achieves a claimed level of privacy.
In this talk we will explore a few techniques in the literature of `estimating’ or `auditing’ DP. I will discuss the technical challenges, as well as new and ongoing work, in extending these techniques to estimate and audit f-DP.