An Equivalence Between Private Classification and Online Prediction
This resolves a foundational theoretical problem in machine learning, connecting privacy and online learning, with implications for boosting privacy and accuracy in learners.
The paper proves that concept classes with finite Littlestone dimension can be learned by approximate differentially-private algorithms, answering an open question and establishing an equivalence between online learnability and private PAC learnability.
We prove that every concept class with finite Littlestone dimension can be learned by an (approximate) differentially-private algorithm. This answers an open question of Alon et al. (STOC 2019) who proved the converse statement (this question was also asked by Neel et al.~(FOCS 2019)). Together these two results yield an equivalence between online learnability and private PAC learnability. We introduce a new notion of algorithmic stability called "global stability" which is essential to our proof and may be of independent interest. We also discuss an application of our results to boosting the privacy and accuracy parameters of differentially-private learners.