Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Locally Optimal Private Sampling: Beyond the Global Minimax

Authors: Hrad Ghoukasian, Bonwoo Lee, Shahab Asoodeh

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We numerically validate our local minimax samplers, showing they consistently and often substantially outperform the global minimax samplers. Figure 1 illustrates that our samplers yield distributions significantly closer to the original under both pure LDP and GLDP. In this section, we numerically compare the worst-case f-divergence of our locally minimax sampler against the globally optimal sampler of [17] under the ̕-LDP setting. Experiments span both finite and continuous domains, evaluating KL divergence, total variation distance, and squared Hellinger distance across ̕ {0.1, 0.5, 1, 2}.
Researcher Affiliation Academia Hrad Ghoukasian Department of Computing and Software Mc Master University EMAIL Bonwoo Lee Department of Mathematical Sciences Korea Advanced Institute of Science & Technology EMAIL Shahab Asoodeh Department of Computing and Software Mc Master University EMAIL
Pseudocode No The paper describes samplers using mathematical formulations and theorems (e.g., Theorem 3.4, Theorem 5.1), but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes Experiment code is publicly available at https://github.com/hradghoukasian/private_sampling, with reproduction instructions provided in Appendix F. For completeness, the code is also included in the supplementary material.
Open Datasets No In the continuous setting with X = R, we fix the universe Plocal and evaluate the empirical worst-case f-divergence over 100 randomly generated client distributions P1, . . . , P100 Plocal. Each Pj represents a client and is constructed as a mixture of a random number of one dimensional Laplace components with scale parameter b = 1; the complete procedure for generating these distributions is described in detail in Appendix E.4.
Dataset Splits No The paper generates synthetic data for its experiments ('100 randomly generated client distributions') rather than using a pre-existing dataset with defined training/test/validation splits. Therefore, it does not specify dataset splits.
Hardware Specification Yes All experiments were conducted on a system running Ubuntu 22.04.4 LTS, equipped with an Intel(R) Xeon(R) CPU @ 2.20GHz and 16GB of RAM.
Software Dependencies No The paper mentions the operating system 'Ubuntu 22.04.4 LTS', but does not provide specific version numbers for any programming languages or libraries used, which are key software components for reproducibility.
Experiment Setup Yes In this section, we numerically compare the worst-case f-divergence of our locally minimax sampler against the globally optimal sampler of [17] under the ̕-LDP setting. Experiments span both finite and continuous domains, evaluating KL divergence, total variation distance, and squared Hellinger distance across ̕ {0.1, 0.5, 1, 2}. ... For the selected values k {10, 20, 100} ... We define the local and global universes as Plocal = P1/3, 3, h L and Pglobal = P1/9, 9, h L, where h L denotes the density of the Laplace distribution with mean zero and scale b = 1. ... In this experiment, to maintain consistency with Park et al. [17], we use K = 10 and k0 = 2.