Research On Key Technologies Of Federated Learning Based On Blockchain

With the rapid development of science and technology and network,the distribution of mobile devices and edge nodes is spreading rapidly,which leads to the extensive distribution of data and computing power.Although machine learning can excavate the value of massive data,privacy protection and network cost lead to data isolation.Therefore,the massive data can not be made full use of.The Federated learning solves the problem of data isolation in technology,but new technology also brings new problems.The problems of model storage,data security,node verification and resource allocation are worthy of further study.However,blockchain based federated learning is not a simple combination of the two.There are some problems in the security verification and incentive methods of blockchain based federated learning.Therefore,in the framework of blockchain based federated learning,this paper optimizes and studies the two key technologies of node verification and node incentive,which is of great significance to promote the popularization and development of Federated learning.Firstly,an optimization method of block chain architecture for federated learning is proposed to solve problem of increasing communication cost caused by node security verification in the Federated learning based on block chain.A two-level aggregation model based on security evaluation in federated learning is proposed by studying the verification mechanism for useless or malicious nodes combining competitive voting verification method and aggregation algorithm.The verified model can effectively reduce the communication cost of node verification in the Federated learning based on block chain by comparing methods in this chapter and the aggregation effect,aggregation speed and communication cost index of the competitive model updating algorithm.Secondly,this paper proposes a federated learning node contribution evaluation method oriented dynamic change in order to solve the problem that the low data quality,reduced participation and stability of edge nodes has a negative impact on the training effect in the federated learning based on block chain.A contribution index model based on node learning quality and subjective and objective evaluation is proposed by studying the profit distribution and incentive methods of each edge node of Federated learning based on block chain.The proposed model can be verified that there is an effective improvement in the incentive effect for high-quality nodes by comparing the training time,cosine distance and Euclidean distance of this method,extended TMC Shapley algorithm and single round reconstruction algorithm,as well as the profits of different types of nodes.