In a massive IoT systems, large amount of data are collected and stored in clouds, edge devices, and terminals, but the data are mostly isolated. For many new demands of various intelligent applications, self-organized collaborated learning on those data to achieve group decisions has been a new trend. However, in order to reach the goal of group decisions, trust problems on data fusion and model fusion should be solved since the participants may not be trusted. We propose a consistent and trust fusion method with the consortium chain to reach a consensus, and complete the self-organized trusted decentralized collaborated learning. In each consensus process, consensus candidates check others’ trust levels to ensure that they tends to fuse consensus with users with high trust, where the trust levels are evaluated by scores according to their historical behaviors in the past consensus process and stored in the public ledger of blockchain. A trust rewards and punishments method is designed to realize trust incentive consensus, the candidates with higher trust levels have more rights and reputation in the consensus. Simulation results and security analysis show that the method can effectively defend malicious users and data, improve the trust perception performance of the whole federated learning network, and make the federated learning more trusted and stable.

在庞大的物联网系统中，大量数据被收集并存储在云，边缘设备和终端中，但这些数据大多是孤立的。对于各种智能应用程序的许多新需求，基于这些数据的自组织协作学习以实现团队决策已成为一种新趋势。但是，为了达到小组决策的目的，由于参与者可能不被信任，因此应该解决有关数据融合和模型融合的信任问题。我们提出了一种与财团链相一致和信任的融合方法，以达成共识，并完成自组织的受信任的分散式协作学习。在每个共识过程中，共识候选人都会检查其他人的信任度，以确保他们倾向于将共识与高度信任的用户融合在一起，信任级别根据分数在过去的共识过程中根据其历史行为进行评估，并存储在区块链的公共分类账中。为了实现信任激励共识，设计了一种信任奖惩方法，信任水平较高的候选人在共识中具有更多的权利和声誉。仿真结果和安全性分析表明，该方法可以有效防御恶意用户和数据，提高整个联合学习网络的信任感知性能，使联合学习更加可信，稳定。信任度较高的候选人在共识中拥有更多的权利和声誉。仿真结果和安全性分析表明，该方法可以有效防御恶意用户和数据，提高整个联合学习网络的信任感知性能，使联合学习更加可信，稳定。信任度较高的候选人在共识中拥有更多的权利和声誉。仿真结果和安全性分析表明，该方法可以有效防御恶意用户和数据，提高整个联合学习网络的信任感知性能，使联合学习更加可信，稳定。