This paper studies joint power control and spectrum access in dynamic flying ad hoc networks (FANETs) under co-channel interference and adaptive jamming. We propose a graph neural network-aided multi-agent reinforcement learning (GNN-MARL) framework that enables distributed, topology-aware decision-making. Each UAV exploits local observations while leveraging graph embeddings to capture time-varying network structure. Simulation results show that the proposed approach improves average throughput by 25-35% compared to DQN-based baselines, while increasing the probability of satisfying QoS constraints by approximately 20%. The method also enhances energy efficiency and maintains fairness among UAVs. Under high mobility and adaptive jamming, performance degradation is limited to within 10%, demonstrating strong robustness. Ablation analysis indicates that removing the GNN module reduces performance by up to 18%, and single-agent variants incur an additional 20-25% loss, highlighting the importance of topology-aware multi-agent learning. Moreover, the framework scales effectively with network size, exhibiting only 5-8% variation in performance. These results confirm that the proposed GNN-MARL framework provides a scalable and robust solution for resource optimization in highly dynamic and adversarial FANET environments.