Carlini-Wagner attack (http://arxiv.org/abs/1608.04644)
To ensure the modifitacions yields to a valid image, we use the third method of box contraints at the paper and we define
and we optimize over the variable
Given x, and choosing a target class
with
To control the confidence with which missclasifation ocurrs we adjust Îș but we wil set it to 0 for out attack
$$||x||{2} = \sqrt{\sum{i=1}^{n} x^{2}_{i}} $$