A communication-efficient semi-supervised federated intrusion detection method for dynamic UAV swarm networks

Dynamic Unmanned Aerial Vehicle (UAV)-oriented federated environments face limited onboard resources, unstable communication links, and heterogeneous local data, making accurate intrusion detection under constrained communication challenging. To address this problem, this paper proposes SSFL-IDS, a semi-supervised federated intrusion detection framework for dynamic UAV-oriented federated environments. Under scarce labeled data and a shared unlabeled data pool, the framework enables distributed intrusion detection through federated learning. To improve collaborative supervision reliability, SSFL-IDS employs a dual-gate filtering mechanism that combines local confidence screening with global vote-consistency verification, so that noisy pseudo-labels can be filtered before propagating across clients. To reduce client-to-server uplink overhead while preserving collaborative distillation, an entropy-guided selective soft-logits upload strategy is further introduced to transmit soft information only for informative unlabeled samples. Dynamic participation and client dropout are also incorporated into the experimental protocol to evaluate robustness under unstable federated environments. Experiments on N-BaIoT, UAVIDS-2025, and UNSW-NB15 show that SSFL-IDS achieves a favorable balance between detection performance and communication efficiency. In particular, it delivers the strongest overall performance on the primary benchmark N-BaIoT, while maintaining a favorable performance–communication trade-off on UAVIDS-2025 and remaining highly competitive on UNSW-NB15 with lower measured client-to-server uplink payload. These results indicate that SSFL-IDS provides an effective and lightweight collaborative solution for intrusion detection in dynamic UAV-oriented and resource-constrained federated environments.