What we cover with iterative Fast Gradient Sign Method (iterative-FGSM) that we can't cover with simple Fast Gradient Sign Method (FGSM)

Asked Jan 14 '20 at 14:42

Active Jan 14 '20 at 14:42

Viewed 288 times

FGSM and iterative-FGSM are two methods of generating adversarial samples. After visiting hundreds of sites and blogs to know that why in iterative-FGSM we perform multiple steps. And what we cover with multiple steps in (iterative-FGSM) that we can't cover in simple FGSM.

But all gives me this unsatisfied answer that in iterative-FGSM we perform FGSM multiple time. Please can anyone gives me satisfied answer to above bold question. Thanks

asked Jan 14 '20 at 14:42

mujeeb

@IanGoodfellow May I have your wisdom.? – mujeeb May 05 '20 at 09:00
As I understand, iterative-FGSM or I-FGSM will recalculate the gradient at each time-step. This is not the same as performing FGSM multiple times since we are changing the gradient. By recalculating the gradient it will be able to find a local optima with less perturbation. – Kevin Apr 24 '21 at 16:23

What we cover with iterative Fast Gradient Sign Method (iterative-FGSM) that we can't cover with simple Fast Gradient Sign Method (FGSM)

0 Answers0