Securing Distributed Gradient Descent in High Dimensional Statistical Learning