Privacy Budget Scheduling
Talkat 15th USENIX Symposium on Operating Systems Design and Implementation (OSDI '21) · 2021-07-14..07-16·
summary
Built an extension to Kubernetes called PrivateKube that treats differential privacy as a non-replenishable resource and implements a scheduler DPF (Dominant Private Block Fairness). DPF unlocks 1/n of each privacy block per arrival, runs a DRF-like allocation on unlocked budget, and in microbenchmarks and an Amazon Reviews NLP workload locates more models than FIFO or round robin while trading higher scheduling delay for up to significantly more allocations; shows particular gains under Rényi DP.
AI-written summary · about
chapters (15)
- 0:00Privacy Budget Scheduling
- 0:25Example: Messaging App
- 1:24What Can Leak?
- 1:53Is Differential Privacy (DP) the Solution?
- 2:32Problem: Privacy Loss Accumulates
- 3:44Solution: DP at Workload Level
- 4:34PrivateKube
- 5:06Architecture
- 6:39Outline
- 7:34Problem: Privacy is not Replenishable
- 8:23Dominant Privacy Fairness (DPF)
- 9:12DPF Example
- 9:54DPF Properties
- 10:20Methodology Questions How does DPF compare to baseline schedulers? How does the DP semantic impact OPF?
- 13:15Conclusion
related talks
45mMaking Privacy Technology Accessible: Benchmarks and Platforms
19mPractical Considerations for Differential Privacy
35mPrivacy amplification by random allocation
10mFair Scheduling for Deep Learning Workloads in Kubernetes
42mDeep Dive Into the Latest Kubernetes Scheduler Features
16mBalancing Privacy: Leveraging Privacy Budgets for Privacy-Enhancing Technologies
26mAaron Roth, Associate Professor, University of Pennsylvania, NYC 2017
20mShare data without revealing personal information