Shrinkwrap: Differentially-Private Query Processing in Private Data Federations
summary
Shrinkwrap implements SMC-based query execution over horizontally partitioned hospital databases and adds differentially-private resizing to replace worst-case padding, plus a cost-model budget splitting optimizer that allocates epsilon across operators. It runs on HealthLink data (15 GB real, scaled to 750 GB synthetic), uses epsilon≈0.5 (delta=5e-5), and swaps to faster SMC backends when available. Results show up to three orders of magnitude speedup over pure oblivious execution and about a 3× gain when switching crypto protocols.
AI-written summary · about
chapters (19)
- 0:00Intro
- 2:41Data Solutions... and Issues
- 8:10Differential Privacy
- 9:06Secure Multiparty Computation
- 10:33Solution Overview
- 18:26Technical Contributions
- 21:06Shrinkwrap Query Processing
- 22:11Privacy Goals and Assumptions
- 26:32Differentially
- 29:15Cost Modeling
- 30:22Privacy Budget Splitting
- 31:16Experimental Results
- 32:51Query Workload
- 33:08End to End Performance
- 34:06Trading Privacy, Accuracy, and Performance
- 35:31Budget Splitting Performance
- 36:48Scaling with Complexity
- 37:07Scaling with Data Size
- 37:41Open Problems
related talks
27mMetadata-private data transfer
45mMaking Privacy Technology Accessible: Benchmarks and Platforms
01h 19mWhen Data Can’t Move but Insight Must
45mPractical Experience with Making Synthetic Data Differentially Private
30mSharing without Sharing: the Case for Data Clean Rooms Powered by Confidential Computing
46mFrom Health Data to Location Traces: Addressing Spatiotemporal Correlations in Differential Privacy
23mSecretFlow-SPU: A Performant and User-Friendly Framework for Privacy-Preserving...
41mAccelerating privacy-enhancing data processing