Comment by systemfreund

It's not clear from the article whether they train the networks with the same shared-key in every iteration, or if they randomize it. Any info on that?