Research On Federated Learning Based Privacy-Preserving Traffic Flow Prediction Scheme

Traffic flow prediction plays an important role in the success of intelligent transportation system.With the development and maturity of Internet of Vehicles technology,governments and organizations can easily collect a large amount of traffic information through vehicles and transportation infrastructure.These organizations have successfully realized intelligent traffic flow prediction based on the collected data.Federated learning is a training framework in which multiple organizations participate and jointly optimize the model under the scheduling of the central server to achieve decentralized training of local data.Federated learning can make these participating organizations get a better model,but they do not need to share their own local data.It follows the principle of partial data collection and minimizing the scope,and reduces the privacy risk brought by traditional centralized learning.However,the federated learning framework does not provide a formal privacy protection mechanism.Malicious servers can use the model uploaded by the organization to update parameters through a single point of failure attack,threatening the security of the original data.Malicious organizations can affect the accuracy of the global model by uploading false or low-quality model update parameters.These problems may cause the failure of the federated learning task,It may even disclose the private information of local data sets of participating organizations.In view of the above problems,the main research contents of this paper are as follows:1.A federated learning algorithm based on LSTM is proposed.For intelligent traffic scenarios,a federal learning traffic flow prediction scheme is implemented.This paper proposes a federated learning algorithm based on LSTM——FedLSTM,which updates the learning model through a secure parameter aggregation mechanism,without the need for organizations participating in joint modeling to directly share local data sets,which greatly improves the feasibility of joint training.Based on the algorithm,this paper designs and implements a traffic flow prediction scheme.The scheme uses differential privacy technology to protect the sensitive information of a single user,and randomly selects organizations participating in the joint prediction model.This scheme obtained real world data and carried out extensive experiments on this scheme.The experiments show that the MAE of FedLSTM is 11.87%lower than the MAE of the worst case(i.e.SVM model)in this experiment,reaching the predetermined accuracy goal.At the same time,this scheme satisfies the(epsilon,0)-differential privacy,and when epsilonis 0.2,it can not only achieve the expected model accuracy,but also achieve meaningful privacy assurance on this basis.2.Based on blockchain technology,a decentralized federal learning traffic flow prediction scheme is proposed.Aiming at the situation that the central server is not trusted,this paper introduces the blockchain technology to realize a discrete,reliable and safe federated learning traffic flow prediction scheme,and realizes joint modeling without a model coordinator.In the proposed scheme,the model update parameters from distributed vehicles are verified by the miners in the blockchain.For the verified qualified model update data,each node on the blockchain will agree on the reliability of the model update data through a consensus mechanism,and finally generate new blocks for the model update data and store them on the blockchain.In addition,in order to further enhance privacy assurance,Gaussian differential privacy mechanism is adopted in this scheme to protect the location information of vehicle users.The scheme obtained real world data and carried out extensive experiments on the scheme.The experiments show that the MAE of the scheme is 8.52%lower than the worst case in the experiment.In terms of privacy protection,the smaller the privacy parameter in the scheme,the more secure the local differential privacy scheme is.In addition,through simulation experiments,this paper simulates attackers to launch poison attacks on the proposed framework.The experiments show that:with the increase in the number of attackers,the ability of frameworks without blockchain solutions to resist poison attacks will decline,which indicates that programs combined with blockchain can enhance the ability to resist poison attacks.