Research On Vertical Federated Learning Optimization Based On Alliance Chai

Federated learning is a distributed machine learning framework with privacy protection that has received continued attention today when data privacy protection is becoming more important.Vertical federated learning as a feature federated enables model training participants to build more accurate models by combining features from different samples without revealing private data.During the model training process,training participants usually need a third party to assist for public-private key pair generation and computation of intermediate parameters.Third party plays an important role,so the third party in vertical federated learning should be credible and reliable.As a weakly centralized and semi-open blockchain with verifiable,traceable,and tamperproof features,the consortium blockchain can build a bridge for cooperation between different organizations and provide trust guarantee between them.Therefore,consortium blockchain can serve as trusted third party in vertical federated learning to assist participants in model training.In this paper,we propose a federated chain-based vertical federated learning architecture and combine verifiable random functions,secret sharing and other techniques to build a trust and security environment to optimize the model training process of vertical federated learning.The specific work is as follows.(1)Propose a consensus algorithm based on verifiable random functions for consortium blockchain,V-Raft.V-Raft is an improvement of the popular consensus algorithm Raft in consortium blockchain.Firstly,the node selection by voting in Raft algorithm is improved by using the lottery algorithm based on the verifiable random function,which improves the speed and stability of node selection;in addition,the setting of the lottery threshold can control the number of winning nodes,which effectively ensures the smooth node selection while reducing the communication overhead and improving the scalability of the coalition chain;secondly,the consensus committee role is added to V-Raft to improve the speed of node consensus.Secondly,the consensus committee role is added to V-Raft to improve the consensus speed of nodes,and at the same time,if the master node fails,the consensus committee can quickly select a replacement node to ensure the smooth completion of model training;finally,simulation experiments are conducted for the V-Raft algorithm.The experiments compare Raft,V-Raft,and KRaft,and the experimental results show that the performance of V-Raft is higher than the other two consensus algorithms.(2)A consortium blockchain-based vertical federated learning model is proposed,where the consortium blockchain uses the V-Raft consensus algorithm to elect consensus committee nodes and master nodes to assist training participants in vertical federation learning.First,a sample alignment scheme based on the consortium blockchain is designed to quickly calculate the common set of samples of the participants;second,a secret sharing algorithm is used for key distribution during the model training process,which can effectively avoid the failure of the model training parameters to be decrypted if the master node fails.