Optimizing differential privacy in federated analytics under known input distributions