Research On Privacy-preserving,Reliable And Fair Federated Learning Scheme

As a hot spot of artificial intelligence technology,federated learning can solve the problem of “data island”.However,existing federated learning schemes face some challenges,such as how to ensure the privacy of updates from the participants,how to avoid unreliable updates and how to encourage participants to contribute their resources to participate federated learning.To solve these problems,this paper formulates the corresponding design goals and propose a privacy-preserving evaluation mechanism to select reliable participants.Considering that participants cannot unconditionally contribute their resources to participate in federated learning,this paper formulates corresponding design goals and designed a fair incentive mechanism based on reinforcement learning to motivate participants.The main work is as follows:· Proposes a Privacy-Preserving and Reliable Federated Learning Scheme(PPRFLS).For the problem that participants may upload unreliable updates,which affects the performance of the global model of federated learning,a lossless privacy-preserving evaluation mechanism based on the similarity of participants' updates is designed.According to this mechanism,a Privacy-Preserving and Reliable Federated Learning Scheme(PPRFLS)is proposed.The PPRFLS adopts dual servers,which is composed of the aggregation server(AS)and the platform server(PS).The PPRFLS uses the CKKS homomorphic encryption scheme to protect the privacy of the participants' updates,and uses the OPTICS algorithm and the quality-based weighted aggregation algorithm to design the privacy-preserving evaluation mechanism.This mechanism is used to evaluate the quality of the updates,and select reliable updates to obtain a reliable global model according to the evaluation results.Since neither inner nor external attackers can obtain the original updates from the participants,the PPRFLS ensures the privacy of the updates.Through performance evaluation and experimental analysis,it is proved that the PPRFLS achieves the reliability by selecting reliable updates to generate reliable global model.· Proposes a Privacy-Preserving,Reliable and Fair Federated Learning Scheme(PPRFFLS).In reality,when participants train the model locally,they will consume computing resources.Therefore,without reasonable payment,participants will not be willing to contribute their resources to the federated learning task.The PPRFLS do not take this into account.Therefore,on the basis of the PPRFLS,an improved scheme-Privacy-Preserving,Reliable and Fair Federated Learning Scheme(PPRFFLS)is designed.Based on the fair incentive mechanism of deep Q network(DQN),the scheme uses deep convolutional neural network(CNN)to compress the learning state space and estimate the Q value of each payment.Without the local privacy information,the PPRFFLS provides fair payment for each group,which means that participants can not obtain higher payment by cheating the PS,which realizes incentive fairness.Through performance evaluation and experimental analysis,it is proved that the PPRFFLS can ensure the incentive fairness of federated learning,encourage participants to upload reliable model updates,and make the PS obtain higher utility.