Toward provably private learning from federated data