Poission Subsampled Rényi Differential Privacy
Authors: Yuqing Zhu, Yu-Xiang Wang
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we conduct various numerical experiments to illustrate the behaviors of the RDP for subsampled mechanisms and showcasing its usage in moments accountant for composition. We will have three set of experiments. (1) We will just plot our RDP bounds (Theorem 5, Theorem 6) as a function of α. (2) We will compare how close the τ-term approximations approximate the actual bound.(5) (3) We will build our moments accountant and illustrate the stronger composition that we get out of our tight bound. Specifically, for each of the experiments above, we replicate the experimental setup of which takes the base mechanism M to be Gaussian mechanism, Laplace mechanism and Randomized Response mechanism. Their RDP formula are worked out analytically (Mironov, 2017) below: |
| Researcher Affiliation | Academia | UC Santa Barbara, Department of Computer Science. Correspondence to: Yuqing Zhu <yuqingzhu@ucsb.edu>, Yu-Xiang Wang <yuxiangw@cs.ucsb.edu>. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | An open source implementation is available at https: //github.com/yuxiangw/autodp. |
| Open Datasets | No | The paper analyzes theoretical mechanisms (Gaussian, Laplace, Randomized Response) and their RDP properties. It does not use external datasets for training machine learning models in the conventional sense. Therefore, there's no mention of publicly available datasets for training. |
| Dataset Splits | No | The paper focuses on theoretical bounds and numerical evaluations of these bounds. It does not involve training models on datasets with explicit training/validation/test splits. |
| Hardware Specification | No | The paper does not specify any hardware details (e.g., CPU, GPU models, memory) used for running the numerical experiments. |
| Software Dependencies | No | The paper does not list any specific software dependencies with version numbers. |
| Experiment Setup | Yes | We will have two sets of experiments with high noise, high privacy setting σ = 5, b = 2, and p = 0.6 and low noise, low privacy setting using σ = 1, b = 0.5, p = 0.9. These parameters are chosen such that the ϵ-DP or (ϵ, δ)-DP of the base mechanisms are roughly ϵ 0.5 in the high privacy setting or ϵ 2 in the low privacy setting. |