Private Geometric Median
Authors: Mahdi Haghifam, Thomas Steinke, Jonathan Ullman
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our main contribution is a pair of polynomial-time DP algorithms for the task of private GM... We also give a simple numerical experiment on synthetic data as a proof of concept that our algorithm improves over DP-(S)GD, as presdicted by the theory. |
| Researcher Affiliation | Collaboration | Khoury College of Computer Sciences, Northeastern University. Supported by a Khoury College of Computer Sciences Distinguished Postdoctoral Fellowship. m.haghifam@northeastern.edu Google Deep Mind. Khoury College of Computer Sciences, Northeastern University. Supported by NSF awards CNS-2232692 and CNS-2247484. |
| Pseudocode | Yes | Algorithm 1 describes our private algorithm Radius Finder for quantile radius estimation... Algorithm 2 Localizationn... Algorithm 3 Loc DPGDn... Algorithm 4 Loc DPCutting Planen... Algorithm 5 SInv Sn... Algorithm 6 DPGD... Algorithm 7 Above Threshold |
| Open Source Code | Yes | In Appendix H, we discussed all the details behind our implementation. Also, we release the code. Since the dataset considered is synthetic, there is no concern regarding the dataset. |
| Open Datasets | No | The dataset consists of two subsets: one tightly clustered at a random location on Bd(R), and the other uniformly distributed over Bd(R). We plot F(θ; X(n))/F(θ ; X(n)) for both algorithms as R varies. The data generation process for this synthetic dataset is described, but there is no explicit link or citation provided for public access to this specific dataset. |
| Dataset Splits | No | The paper describes the use of a synthetic dataset for numerical comparison but does not explicitly mention or specify training, validation, or test dataset splits by percentages, counts, or predefined configurations. |
| Hardware Specification | No | Our results can be produced using public Google Colab. While Google Colab is a compute environment, the paper does not specify the exact GPU/CPU models, processor types, or memory amounts used for the experiments. |
| Software Dependencies | No | The paper mentions the implementation of algorithms and providing code, but it does not specify any particular software dependencies with version numbers (e.g., Python, PyTorch, specific libraries and their versions) needed to replicate the experiment. |
| Experiment Setup | Yes | Hyperparameters. We set the discretization parameter to r = 0.05 in Algorithm 2 and failure probability to 5%. Additionally, we repeat each algorithm 10 times and report the mean. For the other hyperparameters, we used exactly the same hyperparameters as stated in Algorithm 3. For DPGD, we use the hyperparameters in Lemma A.1, and in particular, we choose T such that... |