Research On Lightweight Federated Learning Based On CDH Problem To Protect Data Privacy And Ensure Model Accuracy

Federated learning supports users to use data to train the model locally,and the federated learning server aggregates the local model parameters.But relevant research shows that the local gradient parameters or model parameters uploaded by users have the risk of privacy leakage.Related work proposes privacy protection technologies such as differential privacy,secure multiparty computing,homomorphic encryption,etc.However,secure multiparty computation and homomorphic encryption generally require a lot of communication cost and computation cost,and the communication channel that transmits the key is required to be safe,so it is not suitable for large-scale federated learning scenarios;Differential privacy technology often brings extra noise,which affects the model accuracy and training speed.Therefore,the paper designs a lightweight federal learning framework based on CDH(Computational Diffie-Hellman)to protect data privacy and ensure model accuracy.Firstly,based on CDH problem,a lightweight data encryption protocol is designed by using mathematical modularity.The participating users and the federal learning server generate lightweight keys according to the protocol,and the participating users use the keys to encrypt local model parameters;The federal learning server decrypts the local model parameters of all users with the key,but cannot decrypt the local model parameters of each participating user.Any potential adversary,including semi-honest other users,untrusted servers and malicious third parties,will face CDH problems and cannot decrypt the ciphertext after obtaining the ciphertext of users through insecure communication channels or other means.Therefore,the algorithm in this paper can protect each user's local model parameters from being obtained by other parties except the user,and thus defend against all attacks based on the user's local model parameters.In addition,the data encryption protocol designed in this paper is lightweight,which only generates a small amount of computation overhead and communication overhead,and at the same time can ensure the accuracy of the global model and the training speed.Then,based on the above-mentioned lightweight data encryption protocol,this paper further considers the malicious exit attack and collusion attack of users,and designs an extended parameter encryption algorithm to resist malicious exit attack and collusion attack,so as to ensure that users' local model parameters will not be leaked even if users maliciously quit or collude with each other,and at the same time to ensure the accuracy of the global model.At last,the paper has done a lot of verification experiments on MNIST,Fashion-MNIST and CIFAR-10 datasets,and compared our work with the related work based on differential privacy,homomorphic encryption,and the conventional federated average algorithm(Baseline).A large number of simulation results show that the federated learning framework based on CDH problem designed in this paper can protect the privacy of users' parameters,ensure the accuracy of the model,and generate low computational and communication cost.