Abstract: Existing traffic flow prediction frameworks have already achieved enormous success due to large traffic datasets and capability of deep learning models. However, data privacy and security are always a challenge in every field where data need to be uploaded to the cloud. Federated learning (FL) is an emerging trend for distributed training of data. The primary goal of FL is to train an efficient communication model without compromising data privacy. The traffic data have a robust spatio-temporal correlation, but various approaches proposed earlier have not considered spatial correlation of the traffic data. This paper presents FL-based traffic flow prediction with spatio-temporal correlation. This work uses a differential privacy (DP) scheme for privacy preservation of participant’s data. To the best of our knowledge, this is the first time that FL is used for vehicular traffic prediction while considering the spatio-temporal correlation of traffic data with DP preservation. The proposed framework trains the data locally at the client-side with DP. It then uses the model aggregation mechanism federated graph convolutional network (FedGCN) at the server-side to find the average of locally trained models. The results of the proposed work show that the FedGCN model accurately predicts the traffic. DP scheme at client-side helps clients to set a budget for privacy loss.

Key words: federated learning (FL), traffic flow prediction, data privacy, graph convolutional network (GCN), differential privacy (DP)

摘要： 现有的交通流预测框架由于拥有庞大的交通数据集和深度学习模型的能力，已经取得了巨大的成功。然而，数据隐私和安全性始终是数据需要上传至云端时面临的一大挑战。联邦学习（FL）是一种新兴的分布式数据训练趋势。其主要目标是在不损害数据隐私的前提下训练出高效的通信模型。交通数据具有强大的时空相关性，但先前提出的各种方法并未考虑到交通数据的空间相关性。本文提出了一种基于联邦学习的交通流预测方法，考虑了时空相关性。这项工作采用了差分隐私（DP）方案来保护参与者数据隐私。据我们所知，这是首次在考虑交通数据的时空相关性和差分隐私保护的同时，将联邦学习应用于车辆交通预测。所提出的框架在客户端本地使用差分隐私方案对数据进行训练。然后，在服务器端使用联邦图卷积网络（FedGCN）模型聚合机制来求取本地训练模型的平均值。实验结果表明，FedGCN模型能够准确地预测交通流量。客户端的差分隐私方案有助于用户设置隐私损失预算。

关键词: 联邦学习（FL），交通流量预测，数据隐私，图卷积网络（GCN），差分隐私（DP）