Almost Tight Error Bounds on Differentially Private Continual Counting

The first large-scale deployment of private federated learning uses differentially private counting in the continual release model as a subroutine (Google AI blog titled "Federated Learning with Formal Differential Privacy Guarantees"). In this case, a concrete bound on the error is very relevant to reduce the privacy parameter. The standard mechanism for continual counting is the binary mechanism. We present a novel mechanism and show that its mean squared error is both asymptotically optimal and a factor 10 smaller than …

ai blog blog case concrete continual cs.cr cs.ds cs.lg deployment differential differential privacy error federated learning google privacy reduce release scale standard