In volunteer computing, participants donate computational resources in exchange for credit points. Cheat detection is necessary to prevent dishonest participants from receiving credit points, without actually providing these resources. We suggest a novel, scalable approach for cheat detection in decentralized volunteer computing systems using gossip communication. Each honest participant adapts its detection effort dynamically subject to the number of active participants, which we estimate based on observed system performance. This enables minimizing the detection overhead for each participant, while still achieving a high preselected detection rate for the overall system. Systems based on majority voting usually produce at least $$100\%$$ overhead, whereas our approach, e.g. requires only $$50.6\%$$ overhead in a network with $$1\,000$$ participants to achieve a $$99.9\%$$ detection rate. Since our approach does not require trusted entities or an active cooperation between participants, it is robust even against colluding cheaters.