From Health Data to Location Traces: Addressing Spatiotemporal Correlations in Differential Privacy
Talkat Data Privacy: From Foundations to Applications (Simons Institute workshop, 2019) · 2019-03-04..03-08·
summary
Two projects: building differentially private clinical registries and protecting single-user location traces. For EHRs, tensor factorization runs in a federated way, adds noise to the shared feature matrices with a Gaussian mechanism after proving L2 sensitivity via a Lipschitz bound, and tests on MIMIC-III 40k patients, 200 procedures, 300 diagnoses show noisy phenotypes but nearly identical AUC for mortality prediction at ε≈1; sequential pattern mining fails on 10k ICD‑9 codes unless collapsed to CCS. For mobility, assume a known Markov model, define probable-location sets and a planar isotropic mechanism that beats Laplace for the same privacy budget and gives real‑time event/set protection.
AI-written summary · about
chapters (18)
- 0:00Intro
- 2:47Clinical Data Registry Project
- 4:09EHR Data
- 6:42Efforts
- 7:18Timed Sequences
- 8:34Frequent Patterns
- 9:52Data
- 14:26Spatial Correlations
- 15:52TensorDriven Approach
- 18:05Collaborative Approach
- 19:15Local Factorization
- 20:40Global Optimization
- 24:14Differential Privacy
- 27:30Experimental Results
- 34:00Problem Setting
- 36:57Standard Approach
- 38:04GeoIndistinguishability
- 41:43Spatiotemporal Activities
related talks
42mDifferential Privacy
20mShare data without revealing personal information
15mA Closer Look: Evaluating Location Privacy Empirically
01h 19mWhen Data Can’t Move but Insight Must
01h 12mModelling Privacy Threats of Medical Data and Health Data Spaces
06mPMF: A Privacy-preserving Human Mobility Prediction Framework via Federated Learning
30mAccuracy First: Selecting a Differential Privacy Level for Accuracy-Constrained...
29mPrivacy-preserving methods: Building secure projects