Not All Learnable Distribution Classes Are Privately Learnable

Abstract

We give an example of a class of distributions that is learnable in total variation distance with a finite number of samples, but not learnable under $(\varepsilon, \delta)$-differential privacy. This refutes a conjecture of Ashtiani.

Cite

Text

Bun et al. "Not All Learnable Distribution Classes Are Privately Learnable." Proceedings of The 35th International Conference on Algorithmic Learning Theory, 2024.

Markdown

[Bun et al. "Not All Learnable Distribution Classes Are Privately Learnable." Proceedings of The 35th International Conference on Algorithmic Learning Theory, 2024.](https://mlanthology.org/alt/2024/bun2024alt-all/)

BibTeX

@inproceedings{bun2024alt-all,
  title     = {{Not All Learnable Distribution Classes Are Privately Learnable}},
  author    = {Bun, Mark and Kamath, Gautam and Mouzakis, Argyris and Singhal, Vikrant},
  booktitle = {Proceedings of The 35th International Conference on Algorithmic Learning Theory},
  year      = {2024},
  pages     = {390-401},
  volume    = {237},
  url       = {https://mlanthology.org/alt/2024/bun2024alt-all/}
}