The paper proposes a novel algorithm to hijack the training process so the trained model performs very bad. This is an important topic and all the reviewers agreed that this paper should be accepted.