Abstract
Let 'H be a binary-labeled concept class. We prove that 'H can be PAC learned by an (approximate) differentially private algorithm if and only if it has a finite Littlestone dimension. This implies a qualitative equivalence between online learnability and private PAC learnability.
Original language | English |
---|---|
Article number | 28 |
Journal | Journal of the ACM |
Volume | 69 |
Issue number | 4 |
DOIs | |
State | Published - 16 Aug 2022 |
Keywords
- Differential privacy
- Littlestone dimension
- PAC learning
- online learning
All Science Journal Classification (ASJC) codes
- Software
- Control and Systems Engineering
- Information Systems
- Hardware and Architecture
- Artificial Intelligence