When Less is Enough: Positive and Unlabeled Learning Model for Vulnerability Detection