Mobile crowdsensing possesses excellent collaborative capabilities and has become an important solution for addressing complex sensing tasks. However, the high mobility and anonymity of nodes bring security risks such as forging false data and maliciously consuming platform resources. To address these problems, a cooperative detection scheme for malicious nodes based on Dempster-Shafer(D-S) trust evidence reasoning is proposed. Firstly, the sensing platform collects trust data of nodes through a trust evidence chain, which includes data credibility, recommendation trustworthiness, task cooperativeness and behavioral credibility. Meanwhile, trust decay and control mechanisms are adopted to dynamically update the trust evidence chain. Additionally, a recommendation trust network based on recommendation value is constructed through the collaboration between platform and nodes. The Isolation Forest is used to analyze node behavior, and the behavioral credibility is calculated using Bayes’ rule. Then, K-Means, Gaussian Mixture Model (GMM), Three-way Decision and Grey Cluster are used to classify nodes into trusted, suspicious and malicious sets. Finally, the purification of suspicious set is achieved through raft consensus among trusted nodes. Simulation tests demonstrate the performance of the proposed scheme in accuracy, false negative rate, false positive rate and time consumption. Overall, the scheme has better detection performance when using GMM.