Explore privacy budget scheduling in machine learning through this OSDI '21 conference talk. Delve into the challenges of managing differential privacy (DP) in model training to prevent information leakage. Learn about PrivateKube, an extension to Kubernetes that treats privacy as a manageable resource alongside traditional compute resources. Discover the Dominant Private Block Fairness (DPF) algorithm, designed to handle the non-replenishable nature of privacy budgets. Examine the talk's evaluation of PrivateKube and DPF on microbenchmarks and an ML workload, demonstrating how DPF allows for training more models under the same global privacy guarantee. Gain insights into the complexities of balancing privacy concerns with machine learning model development in this informative presentation.
Overview
Syllabus
Privacy Budget Scheduling
Example: Messaging App
What Can Leak?
Is Differential Privacy (DP) the Solution?
DP at Individual Model Level
Problem: Privacy Loss Accumulates
Solution: DP at Workload Level
Our Vision: Privacy as a Compute Resource
PrivateKube
Architecture
Outline
Problem: Privacy is not Replenishable
Dominant Privacy Fairness (DPF)
DPF Example
DPF Properties
Methodology Questions How does DPF compare to baseline schedulers? How does the DP semantic impact OPF?
Conclusion
Taught by
USENIX
