On the Vulnerability of Graph Learning-based Collaborative Filtering