You've got trust issues, we've got solutions: Differential Privacy
summary
Claim: answer accurate population queries while hiding any single person by adding calibrated noise. Implementation uses the Python Pi DP module to run bounded sum and bounded mean with Laplacian noise (also supports Gaussian and exponential), demonstrates a salary dataset where removing one row no longer reveals an individual salary, and shows Gaussian Naive Bayes training under varied epsilons. Caveats: not for very small datasets or individual-level decisions; typical epsilons cited around 0.01–0.1.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:25Agenda
- 1:17What is privacy
- 2:44Common attacks
- 3:36Case study
- 4:08Voters list
- 4:52Why not anonymize data
- 5:23What is Differential Privacy
- 7:18Differential Privacy Example
- 11:02Local vs Global Differential Privacy
- 12:08Differential Privacy Algorithms
- 13:21Differential Privacy Code
- 16:44Things to consider
- 19:41Trivia
related talks
20mShare data without revealing personal information
35mDifferential Privacy Made Practical
42mDifferential Privacy
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
19mPractical Considerations for Differential Privacy
01h 14mTutorial: Privacy, Part II
44mDifferentially Private Inference for Regression Coefficients
27mPrivacy-preserving statistical analyses