Cognitive Navigation and Resource Allocation for Secure UAV Communications: A Bilevel Formulation and A Hybrid DRL-Convex Optimization Approach

Abstract

Integrating sensing functionality into low-altitude wireless networks (LAWNs) motivates the concept of cognitive navigation and resource allocation (CNRA), where the communication UAV (C-UAV) dynamically designs its navigation and resource allocation policy based on the sensed state of the eaves-dropping UAV (E-UAV) to improve wiretap channel inference and jamming efficiency, thereby enhancing physical-layer security. To further improve secure transmissions throughout the mission duration, this paper formulates a long-term CNRA problem to maximize the number of securely served users over the entire flight period through the joint optimization of UAV navigation, user scheduling, and power allocation. However, the formulated problem is highly challenging due to the strong coupling between long-term UAV navigation and slot-wise power allocation, as well as their heterogeneous decision time scales. To address these challenges, the long-term CNRA problem is reformulated as a bilevel optimization problem, where long-horizon navigation decisions are optimized at the upper level and slot-wise communication and sensing resource allocation are handled at the lower level. Accordingly, a hybrid deep reinforcement learning and convex optimization (DRL-CO) framework is proposed, where soft actor-critic is employed for upper-level navigation learning and convex optimization is adopted for lower-level resource allocation. By exploiting the complementary advantages of data-driven learning and model-based optimization, the proposed method achieves superior secure communication performance, improved training efficiency, and stronger cross-scenario generalization compared with single-level DRL, bilevel DRL, and heuristic navigation baselines.