Differential privacy aims at solving the problem of data anonymization in data analysis, specifically for applications where the background information of an adversary is difficult to estimate. This is often done by adding randomness to the algorithms, which obscures part of the information. The goal is to guarantee that anyone having access to the output cannot with certainty recover individual pieces of information, even if they have additional knowledge about the data set. This usually leads to a trade-off between accuracy and privacy.
The aim of this talk is to:
- Give a motivation and introduction to differential privacy;
- Give a few examples of statistical and machine learning applications that work well with differential privacy;
- Explain the difficulties of applying differentially private methods to queries based on graph structures (like social networks);
- Introduce my research on making online dating recommendation systems differentially private.
The talk is aimed at a broad audience. However, some math is unavoidable.
MetadataTo be recorded?: Yes
URLs for Differential Privacy - An Introduction and an Application
- Friday Aug. 09 15:00 - 16:00 at Speakers Tent