Observable...Yet Still Private? An Offensive Privacy Perspective on Observability
summary
A prototype statistics-only API aims to give analysts trends without raw RUM data while avoiding differential privacy; reconstruction attacks use synthetic-data generation adapted from Amazon research and linear programming with LPsolve in R. Exact LP reconstruction is infeasible at scale; an empirical confidence-rank attack recovers under 2% of rows after ~1,000,000 queries. The chosen mitigation is detection and throttling—API rate limiting and anomaly detection—which forces two-week noisy attacks that are easy to spot.
AI-written summary · about
chapters (13)
- 0:00Introduction
- 0:29What is observability
- 1:32RUM privacy risks
- 2:35Reconstruction attacks
- 3:37Confidence rank explained
- 5:22Malicious user scenario
- 7:47Privacy threat modeling
- 8:52Reconstruction kill chain
- 10:15Attacker setup
- 12:04Confidence rank reconstruction results
- 13:51Mitigating data recovery
- 14:53API rate limiting
- 15:56Conclusion
related talks
40mKeeping data private in real-time pipelines
39mPrivacy-preserving Data Sharing
22mEnhancing Privacy through an Interactive On-demand Incremental Information
32mYou've got trust issues, we've got solutions: Differential Privacy
29mPrivacy-preserving methods: Building secure projects
32mPrivate Insights into AI Use
58mLarge-Scale Private Learning, Part III
12mObservability in Privacy Infrastructure