Quantifying the Preferential Direction of the Model Gradient in Adversarial Training With Projected Gradient Descent

Open in new window