A fundamental problem that confronts decentralized reputation systems is the design of efficient, secure and incentive-compatible mechanisms to gather trust information despite malicious peers, and particularly collusion. This paper presents STORM (for Secure sTructured Overlay for Reputation Management), a P2P protocol which addresses this issue. STORM provides support for gathering peers reputation information safely by self-organizing peers into clusters of peers of common interests. To mitigate the effect of malicious peers, STORM relies on a randomized decision algorithm aiming at securing the routing table maintenance, and on constrained redundant routing focusing on securing information lookup. We show STORM efficiency and robustness through a formal analysis. Specifically, we prove a lower bound on the number of malicious peers that can be inserted in routing tables and we show that our solution is very close to optimal. Finally, we show that our constrained redundant routing approach succeeds at countering collusive behavior.