Federated Learning With Fair Incentives and Robust Aggregation for UAV-Aided Crowdsensing
Yuntao Wang, Zhou Su, Tom H. Luan, Ruidong Li, Kuan Zhang
Abstract
Unmanned aerial vehicles (UAVs) combined with artificial intelligence (AI) have recently gathered significant interest to enable intelligent and on-demand crowdsensing applications. In conventional AI approaches, a wealth of UAVs’ sensory data (which may be privacy-sensitive) needs to be mitigated to the central storage for model training, which poses severe privacy and data misuse risks. The promising federated learning (FL) allows UAVs to cooperatively train a shared model while keeping the private raw data locally. However, FL imposes heavy communication loads on battery-limited UAVs owing to frequent local training and global synchronization. Due to the heterogeneity of UAVs and presence of free-riders and Byzantine UAVs, the quality of UAVs’ model updates can vary dramatically, raising great challenges for fair incentives and robust model aggregation. In this paper, we propose a novel <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">f</u> air and <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">r</u> obust <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">f</u> ederated <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">l</u> earning (FRFL) scheme in UAV-aided crowdsensing. Specifically, we first exploit edge computing-enabled 5G heterogeneous networks in the FL framework to offer proximal FL services with high data rate and low latency for UAVs. Then, based on contract theory, we design an optimal incentive mechanism to precisely and fairly encourage UAVs’ participation in FL under information asymmetry. The method is proved to be truthful, contractual feasible, and computationally efficient. Furthermore, we develop Byzantine-robust aggregation rules and fair model profit allocation rules via contribution index measurement. We also propose a reputation mechanism for credible UAVs recruitment and free-rider prevention by leveraging historical learning records and exploring record freshness for weight assignment. Simulation results validate the efficiency of FRFL in terms of user utility, communication efficiency, and robustness.