English

Not All Learnable Distribution Classes are Privately Learnable

Data Structures and Algorithms 2026-05-20 v4 Cryptography and Security Machine Learning

Abstract

We give an example of a class of distributions that is learnable up to constant error in total variation distance with a finite number of samples, but not learnable under (ε,δ)(\varepsilon, \delta)-differential privacy with the same target error. This weakly refutes a conjecture of Ashtiani.

Cite

@article{arxiv.2402.00267,
  title  = {Not All Learnable Distribution Classes are Privately Learnable},
  author = {Mark Bun and Gautam Kamath and Argyris Mouzakis and Vikrant Singhal},
  journal= {arXiv preprint arXiv:2402.00267},
  year   = {2026}
}

Comments

Appeared in ALT 2024. Fixed a bug and improved exposition. Same version as the one in AM's PhD thesis

R2 v1 2026-06-28T14:33:58.214Z