Privacy-Preserving Incentive Scheme Design for UAV-Enabled Federated Learning

The fusion of federated learning (FL) and unmanned aerial vehicles (UAVs) garnered significant attention as a propitious paradigm, enabling the provision of ubiquitous Artificial Intelligence (AI) services in a privacy-preserving manner. However, despite the intrinsic superiority of FL in safeguarding privacy, an attacker could utilize differential attacks to infer the original data of UAVs. To address the aforementioned challenges, we design a privacy-preserving incentive scheme for UAV-aided FL. In particular, a UAV-aided FL framework is first proposed to facilitate AI model training between UAVs and the server. Then, we quantify the privacy level of UAVs based on differential privacy and analyze its influence on the aggregation accuracy of the server. This scenario involves a complex trade-off between two conflicting objectives. On the one hand, the server desires to obtain higher quality local models for superior aggregation accuracy. On the other hand, UAVs prefer to add more noise to their local models for better privacy protection. Besides, by employing contract theory, we propose an incentive mechanism to optimize the server's aggregation accuracy under information asymmetry. Finally, simulation results validate the superiority and feasibility of our proposed scheme.