Akademik Veri Yönetim Sistemi
AD HOC NETWORKS, cilt.185, 2026 (SCI-Expanded, Scopus)
Flying ad-hoc networks (FANETs) facilitate autonomous communication and collaboration among unmanned aerial vehicles (UAVs) and are increasingly utilized in defense, disaster response, agriculture, and environmental monitoring. However, their limited computational resources and critical operational roles make them susceptible to cyber-physical threats such as jamming, deauthentication, and physical attacks. Existing solutions often target individual attacks and rely on complex, resource-intensive methods that are impractical for lightweight drones. In this study, we propose a novel deep learning-based approach for near-real-time multi-class attack detection in FANETs using RF spectrogram images. RF spectrograms provide a robust, environment-independent representation of drone communications, enabling accurate attack detection without high computational overhead. We introduce DroneAttackRF, the first publicly available real-world dataset of RF spectrograms collected from DJI Ryze Tello and Piranha F-55 drones under various attack scenarios. We develop and evaluate seven deep learning classifiers, including two customized models based on CNN and Autoencoder, as well as five transfer learning models based on VGG-16, ResNet50, InceptionV3, MobileNet, and Xception. The developed models achieved competitive or higher performance compared to prior studies, with the CNN-based model attaining 98.9% accuracy in multi-class detection of different attack types, though dataset and methodology differences limit the feasibility of direct comparison. Additionally, our approach demonstrated fast detection capability, with RF spectrogram acquisition taking only 0.52 s and CNN-based attack classification completing in 0.55 s. The proposed approach demonstrates significant improvements in detection accuracy and efficiency, offering a practical and scalable solution for enhancing UAV network security.