Robust Collaborative Learning with Linear Gradient Overhead