As an emerging sensing data collection paradigm, mobile crowdsensing (MCS) enjoys good scalability and low deployment cost but raises privacy concerns. In this paper, we propose a privacy-preserving MCS system called \textsc{CrowdFL} by seamlessly integrating federated learning (FL) into MCS. At a high level, in order to protect participants' privacy and fully explore participants' computing power, participants in \textsc{CrowdFL} locally process sensing data via FL paradigm and only upload encrypted training models to the server. To this end, we design a secure aggregation algorithm (\texttt{SecAgg}) through the threshold Paillier cryptosystem to aggregate training models in an encrypted form. Also, to stimulate participation, we present a hybrid incentive mechanism combining the reverse Vickrey auction and posted pricing mechanism, which is proved to be truthful and fail. Results of theoretical analysis and experimental evaluation on a practical MCS scenario (human activity recognition) show that \textsc{CrowdFL} is effective in protecting participants' privacy and is efficient in operations. In contrast to existing solutions, \textsc{CrowdFL} is 3 <formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex>$\times$</tex></formula> faster in model decryption and improves an order of magnitude in model aggregation.