Research On Decentralized Collaborative Learning Model Based On Blockchai

With the development of Artificial Intelligence and Intelligent Manufacturing,while actively building intelligent industrial platforms to improve production capacity and reduce costs,enterprises are unwilling to share production data related to core technologies and then choose to build computing centers within enterprises to deploy Artificial Intelligence models.Federated Learning algorithm builds a collaborative training platform among multiple enterprises and finds the rules in diverse data by sharing model parameters and information islands formed on the industrial Internet.However,the existing Federated Learning methods rely on a third-party node as the parameter server.This centralized network structure increases the trust cost between enterprises and faces the risk of collapse and downtime of the central node.This thesis studies a decentralized collaborative learning model based on blockchain,which aims to solve the trust and security problems in Federated Learning.The main research contents are as follows: the design of the block data structure is studied,and the Hash function is used to calculate the message digest for the parameter set as the mark of training rounds.At the same time,the integrity verification of data can be realized to reduce the data packet size in the communication process and realize the reliable transmission of data.Build a decentralized network and create an alliance chain based on Hyperledger Fabric.The work involved includes node generation,channel creation,chaincode deployment,etc.It can be used as the underlying network support for data transmission in the collaborative learning process.Study the asynchronous cooperative learning method,cancel the synchronization lock of Federated Learning so that nodes can upload and download the global model parameters in real-time,and reduce the waiting time of each training round of Federated Learning.Finally,the communication performance experiment tests the communication time,data throughput,and TPS of the decentralized network.The results show that the communication performance of the network can meet the data transmission requirements of collaborative learning training.In addition,this thesis also conducts an Artificial Intelligence experimental analysis on the industrial steam volume prediction data set in thermal power generation.Taking the Mean Square Error,Root Mean Square Error,Mean Absolute Error,and R2 score as the evaluation indexes,combined with the absolute convergence time to evaluate the model training process.The experimental results show that this thesis’ asynchronous cooperative learning algorithm can achieve a good convergence effect.Furthermore,the performance on the test set is also better than the single machine single card regression prediction algorithm.The decentralized collaborative learning model studied in this thesis eliminates the premise of mutual trust among the nodes in the Federated Learning cluster and effectively avoids the security problems of the centralized network,which makes it have an extensive application prospect in the fields of industry,medical treatment,and so on.