Joint trajectory and incentive optimization for privacy-preserving UAV crowdsensing via multi-agent federated reinforcement learning

Abstract

UAV-assisted mobile crowdsensing (MCS) presents a promising paradigm for enhancing data collection in smart city environments, but faces a critical systems challenge: the intricate coupling between spatial trajectory planning, economic incentive mechanisms, and information-theoretic privacy guarantees. Traditional approaches addressing these dimensions in isolation often lead to suboptimal performance. In this paper, we propose PRISM, a unified framework designed around a holistic optimization approach with three components. First, PRISM models the strategic interactions among Service Providers, UAVs, and ground devices through a multi-level Stackelberg game, capturing the hierarchical economic dynamics that influence participation decisions. Second, it incorporates a privacy-aware incentive mechanism that explicitly links UAV navigation decisions to privacy constraints, dynamically managing the privacy-utility trade-off based on data quality. Third, to address the resulting multi-objective optimization across distinct decision timescales, we introduce TMFR, a Two-timescale Multi-agent Federated Reinforcement Learning algorithm that enables UAVs to collaboratively learn policies for both spatial navigation and incentive allocation. Experimental evaluations demonstrate that TMFR achieves 30% faster convergence and 20% higher data quality compared to baselines that optimize only subsets of the problem. These results highlight its suitability for next-generation smart city applications.