Along with the trend towards an autonomous transportation system (ATS), the intelligence of personal mobility service (PMS) can be further lifted by sensing traveler's statuses comprehensively, learning behavior patterns accurately, providing travel options appropriately, and giving service responses timely. Such a process relies on a seamless information flow, which shall address data silos caused by laws and regulations about privacy. This paper proposes a federated architecture for PMS, called FPMS, which adopts federated learning, to provide personalized multi-modal options by aggregating personal data in a privacy-preserving way, and utilizing idle resources of personal devices within the service cluster. In general, by analyzing the physical objects involved, functions required, and data processed, a reference architecture of FPMS is designed to guide its construction in ATS effectively and efficiently. Moreover, a performance evaluation between FPMS and conventional centralized PMS is also presented to reveal the advantages of FPMS in saving service costs.