Long-Short History of Gradients Is All You Need: Detecting Malicious and Unreliable Clients in Federated Learning.