Dynamic Efficient Adversarial Training Guided by Gradient Magnitude