Abstract: Consensus mechanism studies how to achieve data consistency in distributed nodes.As the core technology of blockchain, it directly affects the performance of blockchain system. Byzantine fault tolerant consensus mechanism is the mainstream consensus algorithm of consortium chain, but it has the problems of poor security of accounting node selection and high communication complexity. To address the above problems, firstly, a node reputation value evaluation mechanism based on BP neural network is proposed. Through training, the weight of node reputation value evaluation index is optimized to obtain the node reputation value more accurately. The accounting node is selected according to the reputation value, which can improve the consensus efficiency and reduce the risk of malicious nodes become into accounting nodes, accordingly improve the safety of consensus. Secondly, an adaptive tree communication topology is designed. According to the information entropy of node reputation value, the fork degree of the tree structure is adjusted adaptively to reduce the communication complexity in the traditional P2P topology. In addition, the free fork structure increases the scalability of the communication structure and reduces the negative impact of malicious behavior of node. Finally, the effectiveness of this algorithm is verified by a large number of experimental analysis. When the number of nodes in the blockchain network increases to 500, the throughput of the proposed algorithm is three times that of similar methods, and the delay is only 1/7. At the same time, in the case of continuous switching of accounting nodes and node evil, the throughput of this algorithm is stable and can effectively reduce the impact of node evil. The proposed algorithm has high throughput, low delay and high security, and is suitable for consensus in distributed systems.

Key words: consensus mechanism, Byzantine fault tolerance, neural network, reputation mechanism, adaptive communication topology

摘要： 共识机制研究如何在分布式的节点中达成数据的一致性，作为区块链中的核心技术，直接影响着区块链系统的性能。拜占庭容错类共识机制是当前主流的联盟链共识算法，但存在着记账节点选取安全性差和通信的时间复杂度过高的问题。针对上述问题，提出一种基于BP神经网络的节点信誉值评估机制，通过训练优化节点信誉值评价指标的权重，更为精确地得到节点信誉值。根据信誉值来挑选记账节点，提高共识效率和减小恶意节点成为记账节点的风险，提高共识的安全性。设计了一种自适应的树型通信拓扑结构，根据节点信誉值的信息熵自适应调整树型结构的叉度，减小传统P2P拓扑结构中的通信的时间复杂度，自由叉度结构增加了通信结构的可扩展性并减小了节点作恶带来的负面影响。通过大量的实验分析验证了提出算法的有效性，在区块链网络中节点数增加到500时，该算法的吞吐量为相似方法的3倍，时延仅为1/7。同时在记账节点连续切换和节点作恶情况下，该算法的吞吐量稳定，并能有效降低节点作恶带来的影响。算法具有较高的吞吐量、低时延和高安全性，适用于在分布式系统中进行共识。

关键词: 共识机制, 拜占庭容错, 神经网络, 信誉机制, 自适应通信拓扑