Research On Secure Federated Learning Based On Homomorphic Encryption

With the continuous development of big data and artificial intelligence,how to effectively utilize the massive data generated in people’s production and life has become a widespread concern in all sectors of society.However,in the traditional process of training and learning for large-scale data,various types of security incidents that violate personal privacy data occur frequently,making people pay more attention to the security of private data,resulting in the difficulty of data sharing and the formation of data islands.In order to break the barriers of data islands,the federated learning mechanism has emerged as the times require.Although the federated learning mechanism can collaborate to complete model training under the premise that the data set of multiple participants does not leave the local,which solves the security problems faced by data set sharing to some extent,there are still hidden dangers of privacy leakage in federated learning,such as using inference attacks to derive the data set information.Therefore,it is of great significance to research on a secure and efficient protection mechanism for federated learning privacy data.This dissertation proposes a secure federated learning model based on homomorphic encryption to protect the confidentiality of private data in the federated learning mechanism.In addition,this dissertation designs a secure aggregation scheme on the parameter server side to update the global model parameters in a fairer manner and improve the model training efficiency.After each participant performs model training locally to obtain the corresponding model parameters,the evaluation value is obtained through the model quality evaluation operation,and then the homomorphic encryption technology is used to encrypt the model parameters and upload them to the parameter server.The parameter server uses evaluation values and ciphertext model parameters to perform the secure aggregation operation and distribute the aggregation results to each trainer.Finally,this dissertation demonstrates that the secure federated learning model can protect the confidentiality of the federated learning mechanism through security analysis,and verifies that the scheme can obtain a better model by joint training under the premise of protecting the confidentiality of the federated learning through multiple sets of comparative experiments.In this dissertation,a secure aggregate signature scheme based on elliptic curve and certificateless signature mechanism is proposed,and the scheme is applied to the secure federated learning model.A more secure and efficient trusted federated learning model is proposed,which can effectively deal with the attacker’s malicious tampering of the ciphertext model parameter data,and ensure the confidentiality and integrity of the privacy data in the federated learning mechanism,so that the federated learning mechanism can be applied to the application scenarios with high security requirements.In the federated learning mechanism,after obtaining the ciphertext model parameters that have undergone local model training and homomorphic encryption,each trainer signs and sends them to the parameter server through the aggregate signature mechanism designed in this dissertation.The parameter server performs a signature verification operation to verify the integrity of the received data.Only after the signature verification is passed can subsequent security aggregation operations be performed.In this dissertation,a detailed security analysis and performance analysis are carried out for the proposed scheme.It is proved that this scheme can provide reliable confidentiality and integrity for federated learning,and the practicability of this scheme is verified by experiments.