You take the image recognization machine, and you run the output of your own algorithm through it, training your algorithm to value patterns that confuse the first machine.

I wouldn’t trust this to work against updated machine learning algorithms.