A Game-Theoretic Multi-Agent Reinforcement Learning Approach for Secure UAV-Assisted QKD in Vehicular Networks

Quantum key distribution (QKD) enables secure communication for vehicular networks, but its range limitation calls for multi-hop relaying. Unmanned aerial vehicles (UAVs) can serve as flexible QKD relays, yet their selfish behavior and dynamic mobility pose significant challenges. This paper proposes a secure routing framework for UAV-assisted vehicular QKD that jointly addresses incentive compatibility and path reliability. We model the interaction between the source vehicle and UAVs as a Stackelberg game, where the source sets the relay reward and UAVs decide their participation. A smoothed relay path formulation ensures differentiability and equilibrium tractability. To enable decentralized learning of relay strategies, we adopt a multi-agent deep deterministic policy gradient (MADDPG) algorithm. Furthermore, a Bayesian optimization module is integrated to adaptively search for the optimal reward value that maximizes the source vehicle's utility. Simulation results show that the proposed Stackelberg-MADDPG-Bayesian scheme outperforms existing baselines in terms of relay request success rate and secure key generation rate, and remains robust under various UAV selfishness levels and network dynamics.