Designing an attention-based approach for RF fingerprinting in drone detection and classification
- 1Department of Electrical Engineering, Faculty of Electrical and Computer Engineering, Urmia University
- 2Department of Electrical Engineering, College of Engineering, Al-Iraqia University
Abstract
The growth of unmanned aerial vehicles (UAVs) has generated an increasing need to have an intelligent and efficient radio frequency (RF) fingerprinting system to detect, identify, and classify the mode of operation of the UAV. The paper introduces a computationally efficient transformer-based attention architecture for RF signals analysis in real-world scenario correctly. The architecture possesses a two-stage architecture feature extraction pipeline consisting of the following steps: the correlation filtering feature selection, the ANOVA feature selection, and a multi-headed self-attention encoder to obtain spectral sequences. The evaluation of the model on the large DroneRF dataset indicates that it possesses a high-ranking structure in three hierarchical problems, which are binary UAV detection (DD), four-class UAV classification (DC), and ten-class flight mode classification (FMC). Cross-validation accuracy was 100.00, 99.80, and 99.23; the number of parameters was 39,270 to 152,194, and FLOPs was 2.45 M to 9.81 M, which is appropriate in edge devices with dimensions and block depth customized to task demands. For real-world applicability, the framework was evaluated on the challenging VTI_DroneSET_FFT dataset across 2.4 GHz and 5.8 GHz, in both single- and multi-drone scenarios. A weak baseline (68.09% accuracy) was used to start performance with ablation-guided design, which gave 88.10% accuracy in the 5.8 GHz multi-drone task that was the most challenging. The model scored high in all settings: 98.84%, 96.03%, 89.58%, and 88.09%. The analysis of attention showed that drone detection was effective in various RFs, which were clear to be identified and characterized.
Related
- Traffic surveillance video-based a unified road accident detection, classification, and vehicle tracking framework using an optimized spatio-temporal attention mechanism
- Feature-aware attention convolutional neural network for efficient pulmonary fibrosis classification from chest CT images with FPGA deployment
- TiLCA-Net: nuclei segmentation and classification in pathological images via task-specialized attention and efficient decoding
- CSwinFK: a bidirectional cross-attention CNN-Swin Transformer fusion model for fungal keratitis classification in in vivo confocal microscopy images
- Automated corneal endothelial cell detection and segmentation in specular microscopy images using Attention and Faster U-Net
BibTeX
@article{muneer2026designing,
title = {{Designing an attention-based approach for RF fingerprinting in drone detection and classification}},
author = {Ammar Abdulrasool Muneer and Morteza Valizadeh and Alaa Hussein Abdulaal and Mehdi Chehel Amirani},
journal = {Scientific Reports},
year = {2026},
month = jul,
doi = {10.1038/s41598-026-48487-2},
url = {https://doi.org/10.1038/s41598-026-48487-2},
note = {CC BY},
}
RIS
TY - JOUR
TI - Designing an attention-based approach for RF fingerprinting in drone detection and classification
AU - Ammar Abdulrasool Muneer
AU - Morteza Valizadeh
AU - Alaa Hussein Abdulaal
AU - Mehdi Chehel Amirani
PY - 2026
DA - 2026/07/01
JO - Scientific Reports
DO - 10.1038/s41598-026-48487-2
UR - https://doi.org/10.1038/s41598-026-48487-2
AB - The growth of unmanned aerial vehicles (UAVs) has generated an increasing need to have an intelligent and efficient radio frequency (RF) fingerprinting system to detect, identify, and classify the mode of operation of the UAV. The paper introduces a computationally efficient transformer-based attention architecture for RF signals analysis in real-world scenario correctly. The architecture possesses a two-stage architecture feature extraction pipeline consisting of the following steps: the correlation filtering feature selection, the ANOVA feature selection, and a multi-headed self-attention encoder to obtain spectral sequences. The evaluation of the model on the large DroneRF dataset indicates that it possesses a high-ranking structure in three hierarchical problems, which are binary UAV detection (DD), four-class UAV classification (DC), and ten-class flight mode classification (FMC). Cross-validation accuracy was 100.00, 99.80, and 99.23; the number of parameters was 39,270 to 152,194, and FLOPs was 2.45 M to 9.81 M, which is appropriate in edge devices with dimensions and block depth customized to task demands. For real-world applicability, the framework was evaluated on the challenging VTI_DroneSET_FFT dataset across 2.4 GHz and 5.8 GHz, in both single- and multi-drone scenarios. A weak baseline (68.09% accuracy) was used to start performance with ablation-guided design, which gave 88.10% accuracy in the 5.8 GHz multi-drone task that was the most challenging. The model scored high in all settings: 98.84%, 96.03%, 89.58%, and 88.09%. The analysis of attention showed that drone detection was effective in various RFs, which were clear to be identified and characterized.
SN - 2045-2322
ER -