Adversarial attacks with FGSM (Fast Gradient Sign Method)

Adversarial attacks with FGSM (Fast Gradient Sign Method)

Adversarial attacks with FGSM (Fast Gradient Sign Method) – PyImageSearch

“The FGSM exploits the gradients of a neural network to build an adversarial image, similar to what we’ve done in the untargeted adversarial attack and targeted adversarial attack tutorials.

Essentially, FGSM computes the gradients of a loss function (e.g., mean-squared error or categorical cross-entropy) with respect to the input image and then uses the sign of the gradients to create a new image (i.e., the adversarial image) that maximizes the loss…”

Source: www.pyimagesearch.com/2021/03/01/adversarial-attacks-with-fgsm-fast-gradient-sign-method/

Subscribe to our Digest