Privacy-Preserving Collaborative Training Algorithm For STGCN Via Federated Learning

With the rapid development of industries such as big data and intelligent transportation,ride-hailing has become a common mode of transportation for people.As a result,a large amount of taxi order data has been generated by various companies.Modeling based on historical order data and predicting future order demands can provide reference for tasks such as order matching and hotspot area screening,which is of significant practical relevance.In response to this issue,many researchers have proposed relevant algorithms that effectively address practical problems of order demand prediction.However,the existing order demand prediction algorithms still have many shortcomings.This paper proposes solutions and implementations for the following two shortcomings,and verifies them experimentally:Firstly,the order demand algorithms for taxis lack a universal data processing method.In response to this problem,this paper proposes a prediction model for taxi order data based on a type of model in the spatio-temporal graph neural network-the spatio-temporal graph convolutional neural network,and presents a generalized graph structure information construction algorithm based solely on the latitude and longitude coordinate information.Experiments show that compared with traditional machine learning algorithms,this algorithm performs well in the prediction task of taxi order data.Secondly,taxi order data has extremely high privacy and commercial value,and companies cannot get a better model through data sharing,creating the problem of "data islands".In addition,in the existing graph-based models,the construction process of their graph structure still requires processing of the dataset without privacy protection.In response to this issue,this paper designs a collaborative training algorithm for the spatio-temporal graph convolutional neural network,based on federated learning and homomorphic encryption technologies,to share multi-party data in a privacy-safe manner and collaboratively train a more accurate model.Based on federated learning technology,participants can jointly obtain a global model with lower error without revealing local data.Based on homomorphic encryption technology,data encryption can be performed during the construction of graph structure information and federated training,protecting the privacy of data and models for each participant.Experimental validation shows that in the privacy-protected federated learning scenario,the error of the model is approximately 9.2%lower than the model error obtained from centralized training.Therefore,the algorithm proposed in this paper can collaboratively train a global model with lower error while protecting the privacy of the participant data.On this basis,this paper designs a more generalized federated privacy protection system,including client-side modules,server-side modules,communication modules,and basic functional modules,providing privacy and security protection for local data and model parameters in the training of graph-structured related models.