Because if I understtod anything correctly, the methods tries to find the smallest possible changes that causes the network to make the incorrect classofication. And these smallest possible changes just might be very sensitive to whatever small random ripple weight changes a network has. But I am far from neural network expert, so I can't really answer.