In this project, we enabled Kubernetes/Kubeflow to manage differential privacy (DP) budget among jobs in ML workloads. In particular, our contributions are
- Characterize the privacy resource as dynamically-arriving, non-replenishable private blocks,
- Develop a new scheduling algorithm of DPF (Dominant Private block Fairness),
- Study game-theoretical properties of DPF.
My contributions are
- Reframed the problem and enabled to decouple privacy scheduling from compute resource scheduling, proposed the original DPF algorithm, gave initial rigorous proofs of its game theoretical properties.
- Derived a weaker condition for envy-freeness property, proposed alternative DP scheduling algorithms.
- Proposed the key technique to adapt DPF algorithm to Rényi DP. Namely, RDP allocation curve only need to be partially bounded by unlocked RDP budget curve during scheduling.
- Designed, implemented and fine-tuned a discrete-event simulator.
- Algorithm prototyping and microbenchmarking - evaluated various scheduling algorithms via simulation experiment, investigated their tradeoffs.