New Oracle-Efficient Algorithms for Private Synthetic Data Release

Authors: Giuseppe Vietri, Grace Tian, Mark Bun, Thomas Steinke, Steven Wu

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through empirical evaluation, we demonstrate that our methods scale well with both the dimensionality of the data and the number of queries. Compared to the state-of-the-art method High-Dimensional Matrix Mechanism (Mc Kenna et al., 2018), our algorithms provide better accuracy in the large workload and high privacy regime (corresponding to low privacy loss ε). ... In addition to our theoretical guarantees, we perform a basic experimental evaluation of our algorithms.
Researcher Affiliation Collaboration 1Department of Computer Science and Engineering, University of Minnesota 2Harvard University 3Boston University 4IBM Research Almaden.
Pseudocode Yes Algorithm 1 Primal Framework of No-Regret Dynamics ... Algorithm 2 Data player update in FEM ... Algorithm 3 Data player update in sep FEM ... Algorithm 4 Rejection Sampling Dualquery
Open Source Code Yes We made publicly available the see the exact implementations used for these experiments via Git Hub. For HDMM s implementation see https://github. com/ryan112358/private-pgm/blob/master/ examples/hdmm.py and for FEM s implementation see https://github.com/giusevtr/fem.
Open Datasets Yes We evaluate the algorithms presented in this paper on two different datasets: the ADULT dataset from the UCI repository (Dua & Graff, 2017) and the LOANS dataset.
Dataset Splits No The paper mentions using datasets but does not provide specific training, validation, and test splits (e.g., percentages or sample counts).
Hardware Specification Yes We ran the experiments on a machine with a 4-core Opteron processor and 192 Gb of ram.
Software Dependencies No The paper mentions using 'Gurobi solver' but does not provide a specific version number for it or any other software dependency.
Experiment Setup Yes Our first set of experiments (fig. 1) fix the number of queries and evaluate the performance on different privacy levels. ... Our second set of experiments (fig. 2) fix the privacy parameters and evaluates performance on increasing workload size (or the number of marginals). ... Table 2. First FEM hyperparameters for fig. 1. ... Table 3. Second FEM hyperparameters for fig. 2.