Future autonomous vehicles (AVs) will use a variety of sensors that generate a vast amount of data. Naturally, this data not only serves self-driving algorithms; but can also assist other vehicles or the infrastructure in real-time decision-making. Consequently, vehicles shall exchange their measurement data over Vehicle-to-Everything (V2X) technologies. Moreover, predicting the state of the road network might be beneficial too. With such a prediction, we might mitigate road congestion, balance parking lot usage, or optimize the traffic flow. That would decrease transportation costs as well as reduce its environmental impact. In this paper, we propose a federated measurement and learning system that provides real-time data to fellow vehicles over Vehicle-to-Vehicle (V2V) communication while also operating a federated learning (FL) scheme over the Vehicle-to-Network (V2N) link to create a predictive model of the transportation network. As we are yet to have real-world AV data, we model it with a non-IID (independent and identically distributed) dataset to evaluate the capabilities of the proposed system in terms of performance and privacy. Results indicate that the proposed FL scheme improves learning performance and prevents eavesdropping at the aggregator server side.

翻译：未来自动驾驶汽车将配备多种传感器，生成海量数据。这些数据不仅服务于自动驾驶算法，还可辅助其他车辆或基础设施进行实时决策。因此，车辆需通过车联万物（V2X）技术交换其测量数据。此外，预测路网状态亦具有重要价值——此类预测有助于缓解交通拥堵、优化停车场利用率或调节车流，从而降低运输成本及其环境影响。本文提出一种联邦测量与学习系统：通过车对车（V2V）通信向邻近车辆提供实时数据，同时经车对网（V2N）链路运行联邦学习（FL）机制，构建交通网络预测模型。鉴于尚未获取真实自动驾驶数据，我们采用非独立同分布（non-IID）数据集模拟系统性能与隐私保护能力的评估。结果表明，所提联邦学习方案可提升学习性能，并有效防止聚合服务器端的窃听行为。