STFL:Spatio-temporal Federated Learning for Vehicle Trajectory Prediction
Xuehan Zhou, Ruimin Ke, Zhiyong Cui, Qiang Liu, Wen-Xing Qian
Abstract
Vehicle trajectory data is critical in the field of transportation. Its privacy needs to be protected, but not much attention has been paid to this. Federated learning (FL) has emerged as a useful technique to deal with privacy concerns in a distributed learning manner. Regarding large-scale vehicle trajectory data mining in the intelligent transportation systems (ITS) field, spatio-temporal characteristics are helpful to achieving better model performances; but there is a conflict concerning data sharing between privacy protection and the exploration of the spatio-temporal relationship. To better understand this problem, this paper designs a trajectory spatio-temporal prediction method based on FL named STFL. Different FL clients are trained together without sharing raw data while leveraging the spatio-temporal characteristics. In the overall solution, this paper proposes and integrates two different FL methods, i.e., space trajectory FL (s-FedWvg) and time trajectory FL (t-FedWvg) to form STFL. Several physical characteristics are extracted before training, and the weighted average algorithm is used to enhance the training process. Validation and analysis are conducted with the GAIA Open Dataset, demonstrating promising results using FL on vehicle trajectory data mining.