Research On Secure Aggregation Method Based On Homomorphic Encryption In Federated Learning

Federated learning,as a new distributed machine learning technology,has been widely used in various fields.Its design goal is to achieve efficient machine learning among multiple computing nodes based on ensuring data privacy security and legal compliance.However,in the process of learning and training on large-volume data,potential participants are often unwilling to participate in the federal training to prevent their private data from being leaked,resulting in the existence of data in the form of islands.Therefore,improving the ability to protect user privacy and improving the overall training efficiency are relatively concerns about federated learning at present.Homomorphic encryption technology,as a common means of privacy protection,has also been applied to federated learning.However,when existing federated learning schemes utilize homomorphic properties,homomorphic encryption algorithms with single homomorphic properties are usually used to reduce computational overhead,but their universality is poor and in the process of ciphertext aggregation.The server mostly uses a homomorphic encryption algorithm to realize the ciphertext homomorphic addition operation,which cannot meet the computing requirements of complex aggregation schemes in practical application scenarios.In addition,few existing schemes consider unbalanced data quality distribution of participants.However,data with low shared data quality will affect the overall training accuracy and efficiency,resulting in the loss of the practical application value of the global model.To solve the above problems,this paper studies and proposes a new ciphertext inner product calculation scheme HEVP(Homomorphic Encryption Vector Inner Product,HEVP).A HEFL scheme(Homomorphic Encryption Federated Learning)considering data heterogeneity is designed based on the HEVP scheme.(1)The HEVP ciphertext inner product scheme based on the CKKS homomorphism algorithm is designed for the complex application scenarios of existing federated learning schemes.In this scheme,only one plaintext and ciphertext multiplication operation and one ciphertext addition operation are required to complete the vector inner product operation.Compared with other ciphertext inner product calculation processes,the operation of shift addition is omitted and the complexity of ciphertext inner product calculation is reduced.At the same time,a federated learning framework is constructed based on the HEVP scheme.In the local training stage,the HEVP scheme is used to encrypt the loss function,and the ciphertext results are sent to the aggregation server for aggregation,which has higher accuracy and faster convergence speed.(2)The HEFL scheme is designed to solve the problem of unbalanced data sharing among participants.In the scheme,participants’ data quality design quality score is evaluated to ensure that the parameters generated by the global model are mainly based on high-quality users’ parameters.Meanwhile,in the HEFL scheme,the HEVP scheme is used to ensure users’ data privacy.(3)To verify the effectiveness of the proposed scheme,the deep learning model is compared with the existing scheme in terms of functionality,accuracy,computing overhead,and communication overhead,to consider user privacy and training efficiency.This study presents a new inner product cipher scheme design,to a certain extent solving the homomorphic encryption technology application problems in the study.It can adapt to more complex scenarios,mass fraction,introduced at the same time,reduce the poor quality of the data of participants to the overall training effects,user privacy protection of federal study has a certain positive meaning.