Résumé
We find separation rates for testing multinomial or more general discrete distributions under the constraint of a-local differential privacy. We construct efficient randomized algorithms and test procedures, in both the case where only noninteractive privacy mechanisms are allowed and also in the case where all sequentially interactive privacy mechanisms are allowed. The separation rates are faster in the latter case. We prove general information theoretical bounds that allow us to establish the optimality of our algorithms among all pairs of privacy mechanisms and test procedures, in most usual cases. Considered examples include testing uniform, polynomially and exponentially decreasing distributions.
| langue originale | Anglais |
|---|---|
| journal | Advances in Neural Information Processing Systems |
| Volume | 2020-December |
| état | Publié - 1 janv. 2020 |
| Evénement | 34th Conference on Neural Information Processing Systems, NeurIPS 2020 - Virtual, Online Durée: 6 déc. 2020 → 12 déc. 2020 |
Empreinte digitale
Examiner les sujets de recherche de « Locally private non-asymptotic testing of discrete distributions is faster using interactive mechanisms ». Ensemble, ils forment une empreinte digitale unique.Contient cette citation
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver