Simple PGD Attack Demo for MNIST

What is a PGD Attack?

Projected Gradient Descent (PGD) is a powerful adversarial attack that creates examples to fool machine learning models. It works by iteratively perturbing an input in the direction that maximizes the loss function, keeping the perturbation within a specified budget (epsilon).

Attack Configuration

0.2
10

Results

-
Original Class
-
Target Class
-
Attack Success

Original Image

Prediction: -

Adversarial Image

Prediction: -

Perturbation (x5)