Research On Federated Learning-based Privacy Protection Scheme For Vehicle Trajectory Prediction

Trajectory prediction,which is a significant research direction in the intelligent vehicle field,has great significance for the positive safety protection of intelligent vehicles.Through the data-driven and interaction-aware approach,the interaction features between intelligent vehicles can be well captured,and combined with the vehicle history trajectory data to achieve accurate and reliable vehicle trajectory prediction.However,since trajectory data contains large amount of user privacy information,in order to prevent sensitive information leakage,trajectory data will be stored locally,leading to data silos,which obstructs the training of high quality models based on big data.While federated learning,as an emerging distributed learning method which has attracted much attention from experts and scholars at home and abroad,which adopts the pattern of local training and central aggregation,can enable all parties to connect data silos and build an outstanding global model with better performance by exchanging model parameters or intermediate results without exposing the original data.Nevertheless,the standard federal learning framework itself does not provide comprehensive and adequate privacy protection,and still faces the threat of privacy leakage.To address these issues,this thesis investigates a privacy-preserving scheme for vehicle trajectory prediction based on federated learning.The specific work in this thesis is as follows:(1)For the central aggregation-based federation learning approach in the Telematics scenario,a three-layer architecture of vehicle cluster layer,local data processing layer and cloud core network is built.On the premise of ensuring that the original data does not leave the local area,the MEC nodes only upload model parameters to the cloud core network to achieve federation aggregation.Meanwhile,to protect the model privacy in the modelling process,this thesis designs the FAHEFL encrypted federation network algorithm using FAHE homomorphic encryption and implements session key sharing among federation learning participants based on NTRU-type agent re-encryption.The results show that FAHEFL improves the encryption efficiency by about 5 times over the Paillier-based cryptographic federation network algorithm at the same security level,can achieve the same level of prediction accuracy as centralised learning,and effectively improved the model’s generalisation capability.(2)For the decentralized federation learning approach in the Telematics scenario,the multi-server and multi-client network architecture is used to propose a secure multi-party computation method based on secret sharing to protect the model privacy during the federation learning process.Based on the FAHEFL encrypted federation network algorithm,this thesis constructs a homomorphic message verification code with secret key and adopts digital signature technology to implement a verifiable secret sharing scheme to protect model privacy to a greater extent.In addition,this thesis also designed the GSE-FAHE gradient selection algorithm to optimise the communication overhead in the federation learning training process.Security analysis shows that this scheme satisfies authentication security,federated aggregation security,message integrity and non-repudiation.The experimental results show that this scheme can achieve the same level of prediction as centralised learning,and improves the generalisation ability of this model with high modelling efficiency.(3)Based on the above research,a prototype system for federated learning vehicle trajectory prediction supporting privacy protection is built in this thesis which follows the software development process of requirements analysis,system architecture design,business process design and system implementation.