DPSynth: From Research to Production—Engineering Differentially Private Synthetic...
summary
DPSynth is a marginal-based DP library that turns research into a production system: it measures low-dimensional marginals with epsilon-delta guarantees, fits a private probabilistic graphical model using JAX, then samples synthetic rows at scale. Engineering uses PipelineDP plus Apache Beam and Spark to handle millions of rows and hundreds of columns. It offers auto-marginal selection, schema bounds, APIs for cross-column and cross-row constraints, and is already used for test-data generation, log analysis, and federated-analytics reconstruction; the library is open-sourced and deployed in production.
AI-written summary · about
chapters (10)
- 0:00DP-Synth project intro
- 1:55Core concept and principles
- 3:17User journey and schema
- 4:19How DP-Synth works
- 5:44Scalability challenges
- 6:29Handling real-world data
- 8:32Usability and privacy budget
- 9:16Evaluating data quality
- 11:00Real-world use cases
- 12:06Conclusion and future
related talks
10mCompiling Python Programs into Differentially Private Ones
45mPractical Experience with Making Synthetic Data Differentially Private
01h 14mStatistical Inference and Privacy, Part III
48mSynthetic Data as an Enabler for Learning from Decentralized Private Data
44mDifferentially Private Inference for Regression Coefficients
19mPractical Considerations for Differential Privacy
16mCompute Engine Testing with Synthetic Data Generation
41mpMSE Mechanism: Differentially Private Synthetic Data with Maximal Distributional Similarity