多服务器可追责的隐私保护联邦学习方案OA

With the rapid development of artificial intelligence technology,data-driven machine learning models have been widely used in finance,transportation and medical care.However,large-scale data are scattered and stored in different institutions,forming"data islands",which seriously restricts the rapid development of artificial intelligence.As an emerging framework to break through data islands,federated learning allows participants to co-train models under the premise of protecting data privacy.However,the existing federated learning schemes face challenges such as high risk of privacy leakage and low credibility of aggregation results.To solve these problems,this paper designs a multi-server accountable privacy-preserving federated learning scheme(MSAFL),which has the following characteristics:(1)It is a distributed multi-server federated learning architecture,and uses blockchain technology to build a secure and trusted model aggrega-tion platform;(2)The threshold secret sharing mechanism is used to protect data privacy through gradient fragmentation transmission;(3)Linear homomorphic hashing and blockchain technology are used to establish the verifiable aggregation results and the traceability system of malicious behavior.Theoretical analysis shows that MSAFL scheme has verifiability of results,accountability of malicious servers,robustness and security.Experimental results show that on the MNIST benchmark dataset,compared with the traditional federated averaging algorithm(FedAvg),the MSAFL scheme can effec-tively defend against gradient leakage attacks while ensuring the model convergence performance.Further comparison with other related schemes shows that MSAFL achieves a balance between security and computational efficiency.