Design Of Privacy Preserving Federated Learning Scheme Based On Homomorphic Encryption

With the development of big data technology and cloud computing technology,deep learning has been widely used in a wide range of fields,such as image classification,speech recognition and autonomous driving.However,there is a risk of leaking sensitive information of users in the process of training on large-scale data.Therefore,federated learning methods based on distributed datasets have been developed.However,federated learning also has the risk of privacy leakage,which has greatly limited the development of deep learning.Therefore,it is meaningful to design a security and effective privacy-preserving federated learning scheme for research.This dissertation mainly focuses on the above issues as follows:Our first work is to propose a verifiable federated learning privacy-preserving scheme.Most of the existing privacy-preserving federated learning schemes cannot resist collusion attacks and lack a mechanism for participants to verify the server aggregation results,which leads to problems of user privacy leakage and model training inaccuracy.To address these issues,we first redesigned the federated learning scheme using the El Gamal encryption algorithm.The scheme can effectively protect the data privacy of other participants in the case of complicity between certain participants and the server.At the same time,the scheme uses a bilinear aggregation signature to design a verifying mechanism where participants can effectively verify the correctness and integrity of the server's aggregation results,preventing the problem that the server returns incorrect aggregation results to participants to save computational resources,leading to inaccurate model training results.Compared with existing schemes,our scheme achieves more security while guaranteeing the accuracy of model training.In general,our scheme is a more secure and accurate deep learning solution for federated learning participants.Our second work is to conduct a security analysis of the DeepPAR asynchronous deep learning scheme designed by Zhang et al.and an improved scheme for DeepPAR is proposed to address the security issues of the original scheme.Specifically,this paper demonstrates that the private information of participants in the scheme can be stolen by an honest and curious parameter server by analysing both the insecurity of the re-encryption key generation process in the initialisation phase of the original scheme and the possibility of a collusion attack between the parameter server and the participants.In addition,this paper proposes an improved scheme of DeepPAR based on the original scheme,which redesigns the re-encryption key generation method so that the private key information of the participants and the proxy server transmitted during the generation of the re-encryption key will not be leaked,and ensures that the private key of the proxy server cannot be obtained even if the parameter server and the participants collude,which perfectly avoid the privacy leakage problem in the original scheme.Finally,the security of the improved scheme of DeepPAR is analysed and performance tests are conducted to further demonstrate that our scheme is secure and feasible.