Most research on detecting shilling attacks focuses on users' rating behavior but does not consider that attackers may also attack the users' trusting behavior. For example, attackers may give a low score to other users' ratings so that people would think the ratings from the users are not helpful. In this paper, we define the trust shilling attack, propose the behavior features of trust attacks, and present an effective detection method using machine learning methods. The experimental results demonstrate that, based on our proposed behavior features of trust attacks, we can detect trust shilling attacks as well as traditional shilling attacks accurately.