FGSM Attack Step

Problem

One fast gradient sign method step: x_adv = x + epsilon * sign(grad), where sign(0) = 0. Both x and grad are 2D matrices of the same shape.

The signature is fgsm_step(x, grad, epsilon).

Hint

FGSM moves every input in the direction that most increases the loss, by a fixed step.

case 1

in: [[[1,2],[1,-1]],[[1,1],[1,-1]],0.1]

expected: [[1.1,2.1],[1.1,-1.1]]

case 2

in: [[[0],[5]],[[1],[1]],0.5]

expected: [[0.5],[5.5]]

case 3

in: [[[-1,1],[-1,1]],[[-1,1],[-1,1]],1]

expected: [[-2,2],[-2,2]]

case 4

in: [[[1],[0]],[[0],[0]],0.5]

expected: [[1],[0]]

Starter codePython

def fgsm_step(x, grad, epsilon):
    # Your code here
    pass

4 test cases · run your code against them in the editor