A Huber Loss Minimization Approach to Byzantine Robust Federated Learning