A Saddle-Point Dynamical System Approach for Robust Deep Learning