一种支持自适应联邦学习任务的可信公平区块链框架  

A Trustworthy and Fair Blockchain Framework Supporting Adaptive Federated Learning Task

作　　者：张宝晨 黄月 孔兰菊[1,2] 李庆忠 李文全[1] 郭秋曼 Zhang Baochen;Huang Yue;Kong Lanju;Li Qingzhong;Li Wenquan;Guo Qiuman(School of Software,Shandong University,Jinan 250101;Dareway Software Co.,Ltd.,Jinan 250101)

机构地区：[1]山东大学软件学院,济南250101 [2]山大地纬软件股份有限公司,济南250101

出　　处：《计算机研究与发展》2023年第11期2504-2519,共16页Journal of Computer Research and Development

基　　金：国家重点研发计划项目(2021YFF0704102);国家社会科学基金项目(20BJY131);山东省重大科技创新项目(2020CXGC010106,2021CXGC010108);泉城产业领军人才项目;CCF-华为胡杨林基金项目。

摘　　要：共识机制是区块链技术的重要组成部分,但是主流的共识机制尤其是工作量证明共识机制都存在算力过度耗费和吞吐量低等问题.而联邦学习作为一种分布式机器学习方法,学习模型的本地训练和最终的参与方贡献度计算都需要消耗大量算力资源.因此,提出了一种支持自适应联邦学习任务的可信公平区块链框架TFchain,探索如何利用原本共识机制中耗费的大量算力来提高联邦学习的效率.首先,设计了基于区块链和联邦学习的全新共识机制PoTF(proof of trust and fair),该共识机制将区块链的节点设置为联邦学习的参与方,将原本共识机制中用于哈希计算的大量无效算力转移到联邦学习中,进行本地模型的训练和参与方贡献度的评估;其次,在提高区块链交易吞吐量的同时,对联邦学习的参与方进行了合理的贡献度评估和激励;最后,设计了防止节点作恶的算法.实验结果表明,提出的TFchain能够在回收算力的同时有效提升区块链的交易处理性能,对积极参与联邦学习的参与方进行有效正向的激励.Consensus mechanism is an important part of blockchain technology,but the mainstream consensus mechanisms,especially proof-of-work consensus mechanisms,suffer from problems such as wasted computing power and low throughput.Federated learning as a distributed machine learning method,the local training of learning models and the final calculation of participant contributions require a large amount of computing power.Therefore,we propose a trusted and fair blockchain framework,called TFchain,supporting adaptive federated learning tasks to explore how to utilize the wasted arithmetic power in the original consensus mechanism to improve the efficiency of federated learning.First,we design a new consensus mechanism PoTF(proof of trust and fair) based on blockchain and federated learning,which sets the nodes of the blockchain as the participants of federated learning and transfers a large amount of ineffective arithmetic power used in the original consensus mechanism for Hash computation to federated learning for training of local models and evaluation of participants' contributions.Second,while improving the throughput of blockchain transactions,the participants of federated learning are evaluated and incentivized with reasonable contributions.Finally,an algorithm is designed to prevent nodes from being evil.The experimental results show that the TFchain proposed in this paper can effectively improve the transaction processing performance of the blockchain while recycling the arithmetic power,and provide effective positive incentives to the participants who actively participate in federated learning.

关 键 词：区块链 联邦学习 共识算法 奖励分配 贡献计算 激励机制 作恶检测 算力资源回收 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP311[自动化与计算机技术—控制科学与工程]