Pufferfish Privacy Mechanisms for Correlated Data
summary
Proposes Pufferfish privacy with two concrete mechanisms: a general Laplace-style method that sets noise by the infinity-Wasserstein distance between query-result distributions under alternative secrets, and a practical Markov quilt mechanism for Bayesian/Markov-chain data that picks a quilt, computes a max‑influence correction and adds noise proportional to neighbor count divided by (epsilon minus influence). Experiments show MQM adds roughly tenfold less noise than group DP on activity traces, MQM-exact has better accuracy but higher runtime, and MQM composes like differential privacy (parallel max, sequential sum).
AI-written summary · about
chapters (29)
- 0:00Intro
- 0:24Sensitive Data with Correlation
- 0:48Privacy-Protection under Correlation
- 2:12Differential Privacy
- 3:58The Hope for a Better Solution
- 4:21Components of a Pufferfish Framework
- 5:47E-Pufferfish Privacy
- 7:44Algorithms for Pufferfish Privacy
- 8:18Outline
- 8:30Wasserstein Mechanism
- 16:26Computation Efficiency
- 17:17MQM on Markov Chains
- 18:11MQM Intuition
- 20:12Three Steps
- 20:44Formalizing "Separation"
- 21:24"Small Correction Factor" for Remote
- 22:56For Neighbor
- 23:28Putting Together Everything
- 26:05Experiments: Privacy-Utility Trade-off
- 27:03On Synthetic Data - Short Chain
- 30:31On Physical Activity Data
- 32:05On Electricity Power Data
- 33:21Running Time
- 33:51Composition Properties
- 34:27Two Types of Compositions
- 35:08Parallel Composition - MQM
- 35:57Sequential Composition - Negative Result
- 36:46Sequential Composition - MQM
- 37:07Summary
related talks
43mPufferfish Privacy Mechanisms for Correlated Data
29mVerifying Pufferfish Privacy in Hidden Markov Models
56mFormal Methods and Proofs of Privacy Properties, Part II
20mShare data without revealing personal information
01h 10mDifferential Privacy: Fundamentals to Forefront II
43mPrivacy Amplification by Sampling in Practice
01h 07mStatistical Inference and Privacy, Part I
45mPractical Experience with Making Synthetic Data Differentially Private