基于贡献度证明共识机制的去中心化联邦学习框架

在大数据背景下,保证数据可信共享是数据联邦的基本要求.区块链技术代替传统的主从架构,可以提高联邦学习(federated learning,FL)的安全性.然而,现有工作中,模型参数验证与数据持久化所产生的巨大通信成本和存储消耗,已经成为数据联邦中亟待解决的问题.针对上述问题,设计了一种高效的去中心化联邦学习框架(efficient decentralized federated learning framework, EDFL),能够降低存储开销,并显著提升FL的学习效率.首先,提出了一种基于贡献度证明(proof-of-contribution)的共识机制,使得区块生成者的选举基于历史贡献度而不采用竞争机制,从而有效发避免了挖矿过程产生的区块生成延迟,并以异步方式缓解模型参数验证中的阻塞问题;其次,提出了一种角色自适应激励算法,因为该算法基于节点的工作强度和EDFL所分配的角色,所以能够激励合法节点更积极地进行模型训练,并有效地识别出恶意节点;再者,提出一种区块链分区存储策略,使得多重局部修复编码块(local reconstruction code)可被均匀地分布到网络的各个节点...

Decentralized Federated Learning Framework Based on Proof-of-contribution Consensus Mechanism

In the background of big data,ensuring credible data sharing is the basic requirement of data federal.Using Blockchain technology to replace the traditional client-server architecture can improve the security of federated learning (FL,Federated Learning).However,the huge communication cost and storage consumption generated by model parameter validation and data persistence in existing works have become problems that need to be solved urgently in data federal.To tackle these problems,an Efficient Decentralized Federated Learning framework called EDFL is proposed,which can reduce the system overhead and significantly improve the learning efficiency of FL.Firstly,a consensus mechanism based on PoC (Proof-of-Contribution) is proposed where the election of the block generation is based on historical contribution instead of using the competition mechanism,thus it can avoid the latency of the block generation caused by the mining process,and asynchronously alleviate the congestion in the model parameter validation.Secondly,a role-adaptive incentive algorithm is presented.Because the proposed algorithm is based on the work intensity of participating nodes and the role assigned by EDFL,it can motivate legitimate nodes to actively conduct model training and effectively identify malicious nodes.Thirdly,a novel blockchain partition storage strategy is proposed.The proposed strategy enables multiple Local Reconstruction Code chunks to be evenly distributed to nodes in the network,which reduces the local storage consumption and achieves higher efficiency of data recovery.Lastly,the learning efficiency,storage scalability,and security of EDFL are evaluated in real FEMNIST dataset.Experimental results show that EDFL outperforms the state-of-the-art blockchain-based FL framework from the above three aspects.