Let H be a 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 an equivalence between online learnability and private PAC learnability.