Automatic Algorithmic Choice for Differential Privacy
Differential Privacy is a strong privacy guarantee that comes with a cost: noise must be added to every calculation that deals with private
data. Many algorithms are known that can result in much less-noisy outputs in certain situations. However, the problem of determining the
best algorithm is highly data-dependent. We propose automatically exploring different algorithms to save this burden from falling on
programmers and to help beginners gain insight into differential privacy. We describe a programming language, Jostle, that will determine
the best algorithm during runtime by collecting empirical evidence. The evidence is specified with metafeatures--features about the
algorithms that can be both supplied by the programmer and synthesized by Jostle. We find that Jostle is capable of keeping up with the
individual algorithms that it explores, and that synthesized metafeatures provide new insights into differential privacy.