Privacy in Theory, Bugs in Practice: Grey-Box Testing for Differential Privacy Libraries
summary
An open-source framework called DP Recorder wraps mechanisms for unit tests and uses a record-and-replay approach to run a program on a dataset and its neighbor, force identical mechanism outputs, and check declared sensitivities and budget accounting. It finds 13 bugs in 11 open-source DP libraries, mostly in preprocessing (wrong sensitivity, batch-size/unique-count mistakes, noise-scale collapse) and occasional mechanism mis-implementations. It runs fast, catches many practical failures, and misses floating-point reversibility issues.
AI-written summary · about
chapters (9)
- 0:00Differential privacy testing intro
- 1:36How DP works in practice
- 3:24Bugs in DP frameworks
- 6:34Why verification fails
- 7:59Grey-box testing approach
- 10:47Testing privacy mechanisms
- 13:40Privacy tech maturity gaps
- 14:23Limitations and future work
- 16:10Conclusion
related talks
16mTesting Differential Privacy with Dual Interpreters
19mPractical Considerations for Differential Privacy
45mMaking Privacy Technology Accessible: Benchmarks and Platforms
13mEvaluating the Usability of Differential Privacy Tools with Data Practitioners
47mOpen DP: A Proposal for an Open-Source Suite of Differential Privacy Tools
01h 00mPANEL: Statistical Theory, Privacy and Data Analysis
01h 01mFrom Theory to Practice: Crafting Differential Privacy Systems with Haskell
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017