Unmasking Vulnerabilities: Cardinality Sketches under Adaptive Inputs

Authors: Sara Ahmadian, Edith Cohen

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, our attack used only 4k queries with the widely used Hyper Log Log (HLL++) (Flajolet et al., 2007b; Heule et al., 2013) sketch. We conduct an empirical evaluation of our proposed attack on the Hyper Log Log (HLL) sketch (Durand & Flajolet, 2003; Flajolet et al., 2007a) with the HLL++ estimator (Heule et al., 2013).
Researcher Affiliation Collaboration 1Google Research, United States 2Department of Computer Science, Tel Aviv University, Israel.
Pseudocode Yes Algorithm 1: Attack standard estimators. Algorithm 3: Single Batch Attacker. Algorithm 4: Adaptive Attacker.
Open Source Code No The paper mentions utilizing 'the open-source implementation of HLL++ algorithm in github' but does not state that the code developed for this paper is open-source or provide a link to it.
Open Datasets No The paper states: 'To generate the data... we generate random strings using the English alphabet of a fixed length'. It does not refer to a publicly available dataset or provide access information for the data generated for the experiments.
Dataset Splits No The paper does not provide specific dataset split information (e.g., percentages, sample counts, or detailed methodology for splitting into training, validation, or test sets).
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions utilizing 'the open-source implementation of HLL++ algorithm in github' but does not specify its version number or any other software dependencies with version numbers.
Experiment Setup Yes We set the size of our ground set n k to be in this relevant regime. We consider two different error rates, ϵ = 0.1 with corresponding sketch size k = 104 and ϵ = 0.05, with corresponding sketch size k = 416. We use the same ground set comprising of 5000 keys for both sets of experiments. For each sketch size k, we generate a ground set of size n = 10 · 10 log10(k) to ensure that the ground set is larger than sketch size and the Min Hash component of the HLL++ estimator is used.