Research On Decentralized Fedrated Learning Model Based On Blockchain

The basic idea of federated learning is to aggregate the local model parameters of all participants to obtain a global model,which makes the global model more general and has better performance.However,the traditional federated learning suffers from the problem of model uniqueness.All participants train the local model based on the same global model.But data in different scenarios in reality often have different characteristics and different distributions,and there is no universal model that is suitable for learning data in multiple scenarios.Therefore,the traditional federated learning techniques cannot achieve optimal performance in multiple scenarios.To solve the above problem,a decentralized federated learning model is proposed in this paper.Each participant autonomously selects the remaining models that are similar to the local model and aggregates them to obtain a specific model applicable to the local scenario.This paper further analyzes the security threats faced by the model and proposes corresponding security preventive mechanisms: in order to prevent malicious participants from publishing false models,this paper proposes a new blockchain architecture,which adopts Byzantine faulttolerant consensus algorithm based on model accuracy to achieve secure and trustworthy data sharing among participants;At the same time,in order to prevent malicious participants from obtaining part of others' data from model parameters through backpropagation or other ways,this paper proposes a decentralized federation learning model based on differential privacy to guarantee the privacy of relevant data information when participants share model parameters.The main work and innovation points of this paper are as follows:1.To address the problem that the existing federated learning models are unique and cannot be assigned to different participants to match their scenario models,this paper proposes a decentralized federated learning model.The model aggregates model parameters based on model relevance,and the models with higher relevance to the local model will have more weight in the aggregation process.Model parameter aggregation is not performed at the central server,but locally by the participants.Through this mechanism,the model achieves the specificity of participants' local models,and the same global model is no longer shared among different participants,which effectively improves model performance in different scenarios.2.To address the problem that in decentralized federated learning model,there may be malicious participants publishing false and malicious data poisoning the models leading to model training errors,this paper designs a novel blockchain architecture.This architecture achieves secure and trusted sharing of model parameter information among participants and enhances the fault tolerance and attack resistance of the system.The blockchain transactions contain the model parameters of the participants and the model accuracy.Based on this,a new consensus algorithm,"High Performance Byzantine Fault Tolerance Based on Model Accuracy" consensus algorithm,is proposed.Compared with the traditional PBFT consensus algorithm,this consensus algorithm fully considers the connection between model accuracy and participants' trustworthiness,and establishes consensus groups to complete consensus based on model accuracy.At the same time,the algorithm also optimizes the consensus process by establishing a node information table to realize dynamic node joining and simplifying the consensus process to reduce the consensus time and improve the system performance.3.To address the problem that in decentralized federated learning model,there may be malicious participants who obtain part of others' data from the model parameters by backpropagation or other means,this paper proposes a decentralized federation learning model based on differential privacy.Based on the traditional differential privacy,this paper proposes a differential privacy noise addition scheme for decentralized federated learning by combining the characteristics of federated learning and blockchain.The scheme achieves the protection of model parameter data satisfying ?-differential privacy by adding a certain degree of Gaussian noise to the transmitted model parameters,which can prevent attackers from inferring participant information in the reverse direction.