A Privacy-Preserving Incentive Scheme for UAV-Aided Federated Learning: A Contract Method with Prospect Theory

The convergence of unmanned aerial vehicles (UAVs) and federated learning (FL) has emerged as a promising paradigm to facilitate artificial intelligence (AI) services with enhanced privacy preservation. However, notwithstanding the inherent advantages of FL in terms of privacy protection, attackers can still exploit inference attacks to deduce raw data of UAVs. The existing studies predominantly assume FL servers to be fully rational and have access to all privacy preference information of UAVs (i.e., information symmetry scenario), in the design of privacy-preserving incentive schemes. To tackle these challenges, we propose a privacy-preserving incentive scheme for UAV-aided FL in the presence of information asymmetry while considering FL servers exhibit bounded rationality. Specifically, a practical UAV-aided FL framework is first introduced to enable AI model training between UAVs and the FL server with bounded rationality. In addition, based on differential privacy, we quantify the privacy level of UAVs and subsequently analyze its impact on the aggregation accuracy of the server. This scenario entails two conflicting objectives: the server aims for higher-quality local models to achieve better aggregation accuracy, while UAVs prioritize injecting more noise into their local models to enhance privacy protection. To reconcile the conflicting objectives, we develop an incentive mechanism based on contract theory to optimize the server's aggregation accuracy in the presence of information asymmetry. Furthermore, we employ prospect theory (PT) to the above contract to capture biases in the server's subjective decision-making process. Besides, we deduce closedform solutions for optimal contracts under PT and expected utility theory (EUT), where participants are assumed to be fully rational. Finally, simulation results validate the superiority of our proposed scheme in motivating UAVs to share high-quality local models and improving the aggregation accuracy of the server.