Not All Learnable Distribution Classes are Privately Learnable
arXiv:2402.0026789.33 citationsh-index: 35
AI Analysis
For the differential privacy community, this clarifies the limitations of private distribution learning by showing a separation between learnable and privately learnable classes.
The paper provides an example of a distribution class that is learnable with finite samples but not under differential privacy, weakly refuting a conjecture by Ashtiani.
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 $(\varepsilon, δ)$-differential privacy with the same target error. This weakly refutes a conjecture of Ashtiani.