Probabilistically True and Tight Bounds for Robust Deep Neural Network Training