Research And Implementation Of Federated Learning Technology Based On Blockchain

In recent years,with the rapid development of deep learning technology,the data privacy issues brought by artificial intelligence have attracted a lot of attention.How to develop artificial intelligence technology while protecting data privacy has become a challenge.Federated learning based on blockchain technology allows participants to jointly train a global model without sharing their datasets.It has broad application scenarios and significant value in data privacy protection.However,the current federated learning techniques do not effectively address the Non-IID problem,computational heterogeneity,and decentralization,resulting in unsatisfactory performance and privacy protection.This thesis address these problems and propose the following contributions:(1)This thesis propose a federated learning algorithm based on parallel ensemble learning,using a two-level learner structure composed of primary federated learners and secondary federated learners.This approach improves the performance of federated learning.Experimental results show that our algorithm outperforms other methods when using Non-IID partitioning on the MNIST and Cifar-10 datasets.(2)This thesis propose a decentralized federated learning framework based on directed acyclic graph(DAG)consensus mechanism,enabling asynchronous federated learning.Experimental results demonstrate that,under normal conditions,this framework performs better than traditional federated learning.Even in the presence of computational heterogeneity and malicious nodes,the impact on this framework is much smaller compared to traditional federated learning.(3)This thesis combine the federated learning algorithm based on parallel ensemble learning and the decentralized asynchronous federated learning based on DAG.Using blockchain and smart contracts as the foundation,we develop a blockchain-based federated learning system.The system provides functionalities such as task management,federated learning,and user management.When performing an image classification federated learning task,the system achieves a classification accuracy of 98.64%.