The Flajolet-Martin Sketch Itself Preserves Differential Privacy: Private Counting with Minimal Space

Authors: Adam Smith, Shuang Song, Abhradeep Guha Thakurta

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments: One important attribute of our algorithm is that the entire final value of the sketch including all 1/γ2 basic units but not the hash function descriptions is differentially private...In Section 3, we show that with ε = 1.0, all three estimators reach nearly the same relative error as the non-private estimator, which is below 2% using 4096 hash functions.
Researcher Affiliation Collaboration Adam Smith Boston University ads22@bu.edu Shuang Song Google Research, Brain Team shuangsong@google.com Abhradeep Thakurta Google Research, Brain Team athakurta@google.com
Pseudocode Yes Algorithm 1 AFM: Flajolet-Martin (FM) sketch for distinct elements
Open Source Code No The paper does not contain any explicit statements about releasing source code for the described methodology or links to a code repository.
Open Datasets No The paper mentions 'We consider datasets with true cardinality F0(D) ranging approximately from 2^12 to 2^20 ≈ 10^6.' but does not provide concrete access information, specific links, DOIs, repository names, or formal citations for any publicly available or open dataset.
Dataset Splits No The paper describes experiments and evaluation, but it does not specify any training, validation, or test dataset splits (e.g., percentages, sample counts, or references to predefined splits).
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions 'For fair comparison, we implement all our algorithms in C++.' in Appendix E.1, but it does not provide specific ancillary software details such as library or solver names with version numbers.
Experiment Setup Yes We consider datasets with true cardinality F0(D) ranging approximately from 2^12 to 2^20 ≈ 10^6.