基于声誉的分布式联邦学习节点选择算法

由于隐私泄露的风险越来越大,而采集的数据中的通常包含大量隐私信息,使数据的采集者不愿意共享自己的数据,造成“数据孤岛”,联邦学习能够实现数据不离本地的数据共享,但其在多机构数据共享中还存在一些问题,一方面中央服务器集中处理信息造成昂贵的成本,易产生单点故障,另一方面,对于多机构数据共享而言,参与节点中混入恶意节点可能影响训练过程,导致数据隐私泄露,基于上述分析,本文提出了一种将区块链和联邦学习相结合的以实现高效节点选择和通信的新的分布式联邦学习架构,解放中央服务器,实现参与节点直接通信,并在此架构上提出了一种基于信誉的节点选择算法方案（RBLNS）,对参与节点进行筛选,保证参与节点的隐私安全。仿真结果表明,RBLNS能够显着提高模型的实验性能。

RBLNS: A Reputation-based Learning Nodes Selection Algorithm for Decentralized Federated Learning

Due to the increasing risk of privacy leakage, data collectors are reluctant to share their data private data, which leads to result in "data silos". Federated learning enables data sharing without leaving the local area, but there are still some problems. On the one hand, centralized processing of central server suffers from expensive time cost and single points of failure. On the other hand, for multi-institutional data sharing, model training might be affected by participating nodes mixed with malicious nodes, which leads to data privacy leakage. Therefore, in this paper, a new distributed federated learning architecture is proposed to combine blockchain and federated learning for efficient node selection and communication. And it enables direct communication between participation nodes instead relying on central server. Based on the proposed architecture, a reputation-based node selection algorithm scheme (RBLNS) is proposed to screen the participating nodes and ensure the privacy and security of the participating nodes. The experimental results demonstrate that our RBLNS is capable of improving model performance.