Abstract: Reputation mechanisms are a key technique to trust assessment in large-scale decentralized systems. The effectiveness of reputation-based trust management fundamentally relies on the assumption that an entity's future behavior may be predicted based on its past behavior. Though many reputation-based trust schemes have been proposed, they can often be easily manipulated and exploited, since an attacker may adapt its behavior, and make the above assumption invalid. In other words, existing trust schemes are in general only effective when applied to honest players who usually act with certain consistency instead of adversaries who can behave arbitrarily. In this paper, we investigate the modeling of honest entities in decentralized systems. We build a statistical model for the transaction histories of honest players. This statistical model serves as a profiling tool to identify suspicious entities. It is combined with existing trust schemes to ensure that they are applied to entities whose transaction records are consistent with the statistical model. This approach limits the manipulation capability of adversaries, and thus can significantly improve the quality of reputation-based trust assessment.