Research On Model Aggregation Based On Secure Multiparty Computation In Federated Environment

Federated learning is a novel distributed privacy-preserving machine learning solution.Traditional centralized machine learning,restricted by laws and regulations,cannot further collect user privacy data.Federated learning can solve this problem.Federated learning protects data privacy by aggregating models rather than aggregating data.However,the current aggregation methods all have some defects: the direct aggregation method has security risks and will leak sensitive information;the aggregation method based on homomorphic encryption requires huge computing costs;the aggregation method based on secure multi-party computing has poor robustness and high communication costs.This thesis further studies the aggregation methods based on secure multi-party computing,proposes a secure aggregation protocol with double cloud servers,applies it to federated learning,designs a federated learning model with double cloud servers,and evaluates the performance of protocol and model through experiments.The main contents of this thesis are as follows:(1)The Double-Server Security Aggregation(DSSA)is proposed based on the double cloud server architecture.The DSSA protocol adds a mask to the user’s private data and completes the secure aggregation of private data through the cooperation of two servers;it solves the problem of client disconnection based on secret sharing technology;it uses a pseudo-random generator to reduce communication costs.This thesis analyzes the security and complexity of the DSSA protocol,designs experiments to test the correctness and robustness of the DSSA protocol,and compares the communication cost and computational cost between the DSSA protocol and the homomorphic encryption aggregation method.The experimental results show that the DSSA protocol can safely and accurately aggregate the client’s private data;dropouts cannot affect the normal execution of the DSSA protocol,which has good robustness;compared with other security aggregation methods,the DSSA protocol has a lower computational and communication cost.(2)Based on the DSSA protocol proposed in this thesis,a secure,low-cost,and reliable Double-Serve Federated Learning model(DSFL)is designed.The DSFL model aggregates the local models through the DSSA protocol,which reduces the communication cost and protects the security of the local model.The sum of the data set sizes is calculated based on the secret-sharing to safeguard the security of the data set.A "reliability score" is proposed to consider the difference in user equipment,the influence of abnormal equipment on model training is reduced,and the training efficiency and accuracy of the model are improved.In addition,this thesis analyzes the security of the DSFL model and designs experiments to evaluate its performance of the DSFL model.The experimental results show that the performance of the DSFL model is close to the performance of the traditional machine learning model;adding a "reliability score" helps to improve the accuracy of the DSFL model;dropped clients will not affect the DSFL training process,and the DSFL model has good robustness.