| In recent years,with the rapid development of information technology,artificial intelligence represented by deep learning has been applied to many fields such as fault diagnosis,intelligent control,emotion recognition and biological information.During developing a high-performance deep learning model,research engineers generally need to collect a large amount of data for model training.Due to privacy protection and commercial competition concerns,as well as legal and regulatory requirements,data owners are usually unwilling or even not allowed to share their local data resources,which seriously hinders the further development of deep learning.To break down the data barriers between different institutions,federated learning(FL)is proposed as a novel distributed deep learning approach(Ada PMFL),which can coordinate multiple participants to train a deep learning model.In the FL system,participants replace the traditional way of sharing local data by sharing local model parameters,thereby alleviating participants’ privacy concerns.Although the FL has received extensive attention from academia and industry due to its excellent characteristics,it still faces some problems such as local data information leakage,high cost of cryptographic computation,and unreliable participants.Therefore,it is urgent to improve the performance of the FL system while preserving the local data of the participants.In order to design an efficient and reliable federated learning system,starting from the needs of real industrial scenarios,this thesis conducts research on the adaptive adjustment of the model aggregation interval,the unreliable participant detection algorithm and the hybrid privacy-preserving method for the FL.The main contributions of this thesis can be summarized as follows:(1)To solve the problem that it is difficult to set the global model aggregation interval for the FL system under the unbalance data scenarios,the thesis proposes an adaptive privacy-preserving federated learning approach for non-independent identically distributed data.Firstly,the neural network model is taken as an example to explain how the local model parameters shared by participants leak their local data information,and the reasons for the degradation of the global model availability in the FL system are analyzed by mathematical formulas.Based on this,an adaptive model aggregation scheme is proposed in this paper,which can set the mini-batch size for each participant and adaptively adjust the model aggregation interval to reduce computation and communication costs according to participant training process information.To protect the local data resources of participants,a secure data communication scheme based on Paillier homomorphic encryption is designed in,where participants use the paillier encryption scheme to encrypt their local model parameters.Theoretical analyses and experiment results show that the proposed federated learning approach can realize efficient training of the system and consider the global model availability while protecting the local data resources of the participants.(2)In real industrial scenarios,the FL system may contain unreliable participants which hold low-quality data resources.Model parameters uploaded by unreliable participants would degrade the global model availability and even result in the training process failure.To solve the problem,this paper proposes an unreliable participant detection algorithm based on cosine similarity,and develops the privacy-preserving momentum federated learning resisting unreliable participant approach(Detect PMFL).The unreliable participant detection algorithm cooperates with the participants and the cloud server to calculate the reliability of each participant,to reduce the negative effect of unreliable participants to the global model.CKKS homomorphic encryption scheme is used to protect the local data resources of the participants.Theoretical analyses and experimental results show that the proposed Detect PMFL can effectively reduce the negative effect of unreliable participants in the FL system,and protect the participant data resources.(3)In order to enhance the privacy-preserving level of the FL system,this thesis proposes a privacy-enhanced momentum federated learning approach(PEMFL),which introduces the chaotic system theory into the field of the FL.Participants encrypt the model weights by using the hyper-chaotic encryption technology,so that the cloud server cannot obtain the available model parameters,thereby avoiding direct model parameter interaction between participants and the cloud server.Moreover,utilizing differential privacy and momentum gradient descent techniques,the differential privacy momentum gradient descent(DPMGD)method is constructed,which makes PEMFL resist collusion attacks and improves the convergence speed of the model training.Theoretical analyses and experiment results show that PEMFL can effectively enhance the privacy-preserving level and training performance of the FL system. |
