related question: does epsilon in the fg attack actually correspond to the distortion of the image? Or is that just the coefficient of the gradient in each iteration? If so, eps isn't a great measure of the distortion and we should really be using the distance.
related question: does epsilon in the fg attack actually correspond to the distortion of the image? Or is that just the coefficient of the gradient in each iteration? If so, eps isn't a great measure of the distortion and we should really be using the distance.