Mitigating Sybils in Federated Learning Poisoning