In this paper, we address the problem of designing a robust reputation mechanism for peer-to-peer services. The mechanism we propose achieves high robustness against malicious peers (from individual or collusive ones) and provides incentive for participation. We show that the quality of the reputation value of trustworthy and participating peers is always better than the one of cheating and non participating ones. Finally we formally prove that, even when a high fraction of peers of the system exhibits a collusive behavior, a correct peer can still compute an accurate reputation mechanism towards a server, at the expense of a reasonable convergence time.