You've got trust issues, we've got solutions: Differential Privacy
summary
Differential privacy claims it can hide any single person while still answering population questions by adding calibrated noise; the talk shows concrete tools: PyDP (a Python wrapper around Google’s differential privacy library), Laplace/Exponential/Gaussian noise, and APIs like bounded sum and bounded mean with incremental updates. It demonstrates a real linkage failure (Massachusetts voter + medical leak), gives typical epsilon ranges ~0.1–5, and shows noisy results that approximate true aggregates while warning when small datasets or individual-level needs break the approach.
AI-written summary · about
chapters (13)
- 0:00Differential privacy intro
- 1:06Key privacy terms
- 2:12Anonymization fails
- 4:22What is DP
- 5:49Differential privacy paradox explained
- 8:19Noise ensures privacy
- 10:04DP math definition
- 12:57PDP library
- 14:20PDP algorithms
- 17:54Practical DP examples
- 23:11Local vs global DP
- 24:57When not to use DP
- 27:31Q&A performance & Epsilon
related talks
01h 00mDifferential Privacy: Fundamentals to Forefront I
58mLarge-Scale Private Learning, Part III
29mPrivacy-preserving methods: Building secure projects
39mPrivacy-preserving Data Sharing
26mAlgorithms with More Granular Differential Privacy Guarantees
44mDifferentially Private Inference for Regression Coefficients
16mHow to Break, Then Fix, Differential Privacy on Finite Computers
32mBLENDER: Enabling Local Search with a Hybrid Differential Privacy Model