Individualized Privacy Accounting via Subsampling with Applications in Combinatorial Optimization
Abstract
In this work, we give a new technique for analyzing individualized privacy accounting via the following simple observation: if an algorithm is one-sided add-DP, then its subsampled variant satisfies two-sided DP. From this, we obtain several improved algorithms for private combinatorial optimization problems, including decomposable submodular maximization and set cover. Our error guarantees are asymptotically tight and our algorithm satisfies pure-DP while previously known algorithms (Gupta et al., 2010; Chaturvedi et al., 2021) are approximate-DP. We also show an application of our technique beyond combinatorial optimization by giving a pure-DP algorithm for the shifting heavy hitter problem in a stream; previously, only an approximate-DP algorithm was known (Kaplan et al., 2021; Cohen & Lyu, 2023).
Cite
Text
Ghazi et al. "Individualized Privacy Accounting via Subsampling with Applications in Combinatorial Optimization." International Conference on Machine Learning, 2024.Markdown
[Ghazi et al. "Individualized Privacy Accounting via Subsampling with Applications in Combinatorial Optimization." International Conference on Machine Learning, 2024.](https://mlanthology.org/icml/2024/ghazi2024icml-individualized/)BibTeX
@inproceedings{ghazi2024icml-individualized,
title = {{Individualized Privacy Accounting via Subsampling with Applications in Combinatorial Optimization}},
author = {Ghazi, Badih and Kamath, Pritish and Kumar, Ravi and Manurangsi, Pasin and Sealfon, Adam},
booktitle = {International Conference on Machine Learning},
year = {2024},
pages = {15491-15511},
volume = {235},
url = {https://mlanthology.org/icml/2024/ghazi2024icml-individualized/}
}