Smart city mobility increasingly depends on fleets of autonomous and remotely piloted vehicles that support logistics, delivery, and emergency response. Their continuous connectivity through wireless control and sensing expands the attack surface and exposes every node to diverse, rapidly evolving cyber threats. A static detection model cannot remain effective across such heterogeneous and dynamic environments. At the same time, most aerial or ground vehicles operate under strict energy and computation limits, which prevent the use of heavy deep neural networks (DNNs). Intelligent decision-making is also required to coordinate local learning and global updates efficiently. To address these challenges, we design a hierarchical intrusion detection framework that performs federated training at a ground control station (GCS) using lightweight models and non-IID client updates, while UAVs participate as distributed learners without sharing raw traffic data. Each client trains a compact encoder with a scaled cosine classifier head using a hybrid objective that combines supervised learning with an episodic few-shot prototypical regularizer. A deep Q-network (DQN) controller on the ground decides whether to execute or skip a federated round to reduce unnecessary communication and computation once performance stabilizes. This architecture allows rapid local adaptation and global resilience without overloading airborne resources. Experiments on the UAVIDS-2025 dataset show that the proposed hybrid federated learner achieves about 98.5 % detection accuracy in the no-attack setting and about 94 % under byzantine conditions when defenses are enabled, while the DQN-based scheduler can skip selected rounds without changing the underlying training pipeline. These results demonstrate a practical path toward resource-aware federated intrusion detection for smart-city UAV operations.
Islam et al. (Thu,) studied this question.