The Fast Gradient Sign Method (FGSM) generates an adversarial example as $x' = x + \epsilon \cdot \mathrm{sign}(\nabla_x L(\theta, x, y))$. Why use the sign of the gradient rather than the gradient itself?

Adversarial Machine Learning

1 selectedDifficulty 5-51 unseenView topic

IntermediateNew

0 answered

1 intermediateAdapts to your performance

Question 1 of 1

120sintermediate (5/10)conceptual

The Fast Gradient Sign Method (FGSM) generates an adversarial example as x^{'} = x + ϵ \cdot sign (\nabla_{x} L (θ, x, y)) . Why use the sign of the gradient rather than the gradient itself?