Design And Implementation Of A Decentralized Federated Learning And Data Sharing Platform

China is building an international governance system for cyberspace applicable to the next generation of IPv6,which involves multiple autonomous domains(countries,regions,organisations,etc.).Among the IPv6 network active measurement tasks,artificial intelligence models are also being used on a large scale,but there is a risk of cross-border transmission of model training data and measurement task privacy data between multiple autonomous domains.How to achieve privacy data sharing and joint modelling in multi-domain networks while protecting data privacy is one of the important requirements for this task.Federated learning is an effective solution to this problem.Federated learning enables multi-party joint modelling without exposing local training data,but federated learning systems are also vulnerable to model poisoning and model privacy inference attacks.By decentralising federation learning using blockchain technology and encrypting models using differential privacy techniques,the security and data privacy of federation learning systems can be further protected.For joint modelling and privacy data sharing in active measurement tasks in multi-domain networks,this thesis implements a decentralised federation learning and data sharing platform in conjunction with blockchain technology.The main research elements are as follows.(1)This thesis designs and implements a decentralised federal learning framework that incorporates blockchain neighbourhood council election mechanisms and differential privacy algorithms.The federal learning global model maintenance and local model update storage in this framework are implemented through blockchain smart contracts,which can prevent the model from poisoning attacks.Due to the poor performance of the blockchain traditional miner consensus mechanism,the framework adopts the consensus mechanism of blockchain neighborhood council election instead of miner consensus mechanism,which can reduce the blockchain consensus computation and improve the system performance.The framework also adopts a Gaussian differential privacy algorithm based on the federated learning model to further encrypt and prevent model privacy inference attacks.2)This thesis designs and implements a data sharing framework based on blockchain decentralised storage.The framework uses blockchain decentralised storage technology to store privacy data between multi-domain networks;and uses blockchain smart contracts to implement data encryption,data sending and receiving,etc.3)Combining the above framework to realise the entire decentralised federal learning and data sharing platform,the platform adopts a multinode distributed architecture,with single nodes built using front and backend separation technology,and decentralised communication and data storage between nodes is achieved through blockchain technology.The frontend of the node is built with Vue framework and Element UI component library,the back-end is built with Django framework based on Python language,and the local data is stored with Mysql database.The decentralised data storage between nodes is implemented using the Swarm blockchain network,and the federated learning mechanism and data sharing mechanism based on the Habitat mechanism are implemented using the Fisco blockchain network.This thesis first introduces the background of the topic,briefly explains the current status of federated learning and blockchain research at home and abroad,and summarises the key technologies in this thesis.Based on the background,this thesis analyses the system requirements,including functional requirements,non-functional requirements,etc.Based on the requirements analysis,this thesis introduces the overall architecture of the system,which is divided into four layers:presentation layer,access layer,service layer and storage layer;introduces four modules:system management,federated learning,data sharing and blockchain;and describes the overall design of the database using E-R diagrams.Secondly,this thesis details the decentralised federal learning framework that incorporates the blockchain neighbourhood council election mechanism and differential privacy algorithm.Again,this thesis presents the detailed design and implementation of the main modules of the system,and introduces the relevant module implementations through activity diagrams,class diagrams and pseudo-code,including decentralised federation learning,decentralised data sharing,blockchain and backend management modules.Following this,the thesis details the development and deployment environment of the system and tests the functionality and performance of the system.The test results show that using this system for federated learning can achieve model results that approximate traditional training methods while preserving the privacy of federated learning models.The thesis concludes with a summary of the thesis work,a description of the problems to be solved and an outlook.The research designed in this thesis is based on the development of a decentralised federation learning and data sharing platform for network active measurement tasks in the national key project "Next Generation IPv6 International Cyberspace Governance System"(Project No.2020YFE0200500).The research results of this thesis can solve the problem of data privacy protection in multi-domain network active measurement tasks,multi-party federated modelling and data sharing.