Generating realistic, differentially private data sets using GANs
summary
Generates synthetic individual-level data with differential privacy by adding noise to gradients during GAN training and releases a PyTorch implementation called the DPWGAN package. Trains on 2017 ACS PUMS Illinois data (~600,000 records) with four categorical fields (age 9 bins, marriage 5, transit 12, schooling 24), outputs a synthetic dataset the same size, and reproduces cross-tabs and broad correlations while providing formal differential privacy guarantees.
AI-written summary · about
chapters (19)
- 0:00Introduction: Opening
- 0:30Introduction: Motivating problem
- 1:31Introduction: Traditional solutions
- 3:33Introduction: Alternative approach
- 5:13Introduction: Lecture outline
- 5:58Differential privacy
- 13:15Generative Adversarial Networks (GANs)
- 17:03Private models
- 19:20DPWGAN package
- 20:20Example: US Census
- 20:37US Census: Dataset background (2017 ACS PUMS data)
- 21:40US Census: Generating synthetic data
- 22:19US Census: Synthetic data generator architecture
- 23:01US Census: Data generator code
- 23:41US Census: Discriminator architecture
- 24:10US Census: Discriminator code
- 24:26US Census: DPWGAN code
- 24:56US Census: Data comparison
- 28:18Conclusion
related talks
10mCompiling Python Programs into Differentially Private Ones
01h 14mStatistical Inference and Privacy, Part III
30mPrivate Generative Models: Balancing Privacy and Utility in AI
14mEstablishing Privacy Metrics for Genomic Data Analysis
16mGenerating High-Quality Tabular Synthetic Data at Scale
48mSynthetic Data as an Enabler for Learning from Decentralized Private Data
35mSynthetic Data for Localized Solutions
15mDifferentially Private Algorithms for 2020 Decennial Census Detailed DHC Race & Ethnicity