Research On Security Protection Of Centerless Federated Learning Based On Blockchain

In order to solve the data island and data security dilemma faced by machine learning,the concept of Federated learning was proposed for the first time.In recent years,federated learning has also attracted great attention in the field of information security,among which the most concerned problem is the security of data and models in federated learning.The current centralized federated learning scheme is prone to single point of failure and evil doing by the central server,so blockchain is used to replace the central server.However,the blockchain network still cannot avoid the poisoning attack of participants on the global model and the reasoning attack on participants’ local data.At present,good research results have been achieved on the security architecture of centerless federated learning.Most of the existing schemes store the local updates of the federated learning global model and parameters directly on the blockchain,and realize participant authentication through the blockchain,so as to avoid tampering with the global model and local updates.However,the above scheme can not take into account the global model security and participant local data security.To solve the above problems,this dissertation proposes a security protection scheme for centerless federated learning based on blockchain.The main research contents and contributions are as follows:(1)A blockchain based centerless federated learning security framework is proposed to solve the problems of global model security and local data security in the existing solutions.In the overall framework,the design of node trusted supervision mechanism based on credibility,block storage structure and consensus mechanism based on model training quality are designed to realize federal learning security training,and the security and performance of the above framework are analyzed.(2)A specific design and implementation method of node trusted supervision mechanism in the training process of centerless federated learning is proposed.By designing the node credibility evaluation mechanism,the node credibility is evaluated from the two dimensions of accuracy and participation times,so as to improve the fairness of node credibility evaluation,and use Wilson confidence interval to reduce the impact of participation times on node credibility.By designing a verification node election scheme based on node reliability,we can avoid the evil of verification nodes and improve the security of the framework.(3)For the above proposed scheme,this dissertation establishes the FISCO BCOS blockchain environment,deploys the federated learning algorithm,and uses the cat and dog classification data set for experimental verification.In order to evaluate the above scheme,this dissertation reproduces the federal learning scheme based on centralized server and the federal learning scheme based on committee mechanism as a comparative experiment.In the experimental process,firstly,the accuracy of the global model and the training time of the model are verified.After 400 rounds of model training,the highest accuracy of the global model is 95.73%,which is better than that of the scheme in this dissertation;Secondly,in the experiment,malicious participants and malicious analysts are set to verify the security protection ability of the scheme.The experimental results show that the scheme is more secure and can resist less than the collusion of 70% malicious nodes.