projected gradient descent (pgd)

PGD is a method to generate adversarial examples, which are inputs to a neural network that induce it to make a mistake.