RF-STARNet: A Multi-modal Lightweight Fusion Framework of UAV RF Fingerprinting Identification for Public Low-Altitude Safety
Beijing University of Posts & Telecommunications.
- [06/Nov/2025] Initial public release of RF-STARNet repository.
- [20/Nov/2025] Complete the experiments for the proposed innovations, design rigorous comparative experiments, and begin preparing figures for this paper.
- [04/Dec/2025] The main body of the paper is essentially complete. Detailed updates have been made to the GitHub repository.
- [22/Jan/2026] Finish the paper writing and update the GitHub repository.
- [31/Jan/2026] Submit the paper, waiting for review.
- [19/May/2026] Major revisions to the paper.
- [13/Aug/2026] Congratulations! My article has been published in the Early Access area on IEEE Xplore.
With the rapid development of Low-Altitude Wireless Networks (LAWN), the effective management and regulation of Unmanned Aerial Vehicle (UAV) is critical for low-altitude safety. Radio Frequency (RF) fingerprinting is a key technology for UAV detection due to its low power consumption and broad coverage. However, the state-of-the-art UAV RF detection schemes have a lower efficiency because RF signals or spectrograms are generally treated as generic sequences or images, failing to capture inherent stationary features. Consequently, massive dynamic redundant features are over-learned, resulting in limited identification performance and increased model complexity. To address these, we propose a multi-modal lightweight RF Spectro-Temporal Attention Residual Network (RF-STARNet), which treats the signal's frequency and time characteristics as distinct modalities. Specifically, the Spectrum-Aware Residual Network (SARNet) is designed to extract sparse and stationary spectral patterns, while the Sparse Sub-band Temporal Attention (SSTA) module captures global temporal autocorrelation and cyclostationary features. Furthermore, the Adaptive Gating Mechanism (AGM) dynamically calibrates the contributions of each modality branch, enforces feature diversity by a novel deep supervision strategy, and complementary discriminative knowledge is learned therein as a consequence. Experimental results demonstrate that RF-STARNet achieves a superior accuracy of 81.44% with a low computational cost of 3.8M Floating Point Operations (FLOPs), exhibiting remarkable robustness even in low Signal-to-Noise Ratio (SNR) environments.
The overall architecture of RF-STARNet consists of three main stages:
- SARNet (Spectral Branch): Extracts time-averaged PSD features using multi-dilation convolutions to capture micro-scale and macro-scale spectral details.
- SSTA (Temporal Branch): Utilizes local average pooling and grouped self-attention with relative position bias to model long-range temporal dependencies (e.g., OFDM symbol periodicity, FHSS hopping).
- Gated Fusion & Deep Supervision: Uses robust spectral features to guide temporal activation and employs multi-loss optimization (Main + Aux + Orthogonal) to ensure feature complementarity.
We use the DroneRFa dataset for evaluation.
- Signals: Raw I/Q signals collected via USRP.
- Preprocessing: Signals are sliced into 50ms segments and transformed into 512x512 Spectrograms via STFT.
- Categories: 10 classes (9 Commercial UAVs + 1 Background Noise).
Data Structure
Please organize the dataset as follows:
data_set_5e6/
├── T0000/
├── T0001/
├── T0010/
├── ...
└── T1001/- Clone the repository:
git clone https://github.com/lzk2211/RF-STARNet.git
cd RF-STARNet- Install dependencies:
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.4 -c pytorch -c nvidia
pip install -r requirements.txtThe main entry point is main.py. You can switch between training and testing modes using command line arguments.
Training
To train the RF-STARNet (Small version) with Multi-Loss Deep Supervision:
python main.py \
--mode Train \
--model F_T_small_mulitloss\
--exp_name F_T_small_mulitloss \
--root_path /path/to/your/data_set_5e6 \
--batch_size 64 \
--epochs 250 \
--lr 0.0001 \
--seed 42Key Arguments:
--epochs: Number of training epochs--lr: Learning rate--weight_decay: Weight decay for optimizer--batch_size: Batch size--resume: train from checkpoint--ratio: ratio--mode: switch mode--exp_name: the name of the experiment--class_num: class number--model: switch model--root_path: dataset path--seed: Random seed for reproducibility--device: Device to use for training/testing
Performance comparison on DroneRFa dataset (Avg Accuracy vs. Efficiency)
The proposed dual-stream fusion network, RF-STARNet, demonstrates superior performance across all evaluation metrics compared to its individual sub-branches (SARNet and SSTA). specifically, RF-STARNet achieves an average accuracy of 81.44%, outperforming the standalone SARNet and SSTA branches by 3.63% and 5.65%, respectively. This validates the effectiveness of fusing temporal and spectral features, leveraging the complementary strengths of CNN-based and Transformer-based architectures.
(a) Accuracy |
(b) Precision |
(c) Recall |
(d) F1-score |
The trade-off between model accuracy and computational complexity.
Lightweight Design: RF-STARNet maintains an extremely low parameter count (0.188 M) and computational cost (3.80 M MACs). Compared to the heavyweight ResNet18 (11.18 M Params), our model reduces the model size by approximately 98% while achieving a higher accuracy (81.44% vs. 76.11%).
Inference Speed: The optimized architecture allows for rapid inference (3.65s on GPU), making it highly suitable for real-time drone detection on edge devices. The high Model Efficiency Index (MEI = 6.87) further confirms that RF-STARNet achieves the best balance between performance and resource consumption among all compared methods.
(a) The Average Accuracy vs. Computational Cost |
(b) The Average Accuracy vs. Model Size |
Part I: Impact of specific modules on the baseline architecture.
Part II: Benchmarking against comparison methods.
| Model / Configuration | Avg Acc (%) | Params (M) | MACs (M) | MEI (Score) | GPU Infer. Time (ms) | CPU Infer. Time (ms) | Memory (MB) | Train Cost (GPU/h) |
|---|---|---|---|---|---|---|---|---|
| Part I: Module Ablation Study | ||||||||
| 1. Frequency Domain Architecture: SARNet | ||||||||
| Baseline | 69.43 | 0.019 | 0.59 | 208.32 | 0.70 | 0.97 | 0.18 | 0.09 |
| w/ SE block | 70.29 (+0.86%) | 0.031 | 0.68 | 207.84 (-0.23%) | 0.68 | 0.96 | 0.18 | 0.09 |
| w/ MRFSM | 72.53 (+3.10%) | 0.019 | 0.59 | 227.34 (+8.37%) | 0.69 | 1.03 | 0.18 | 0.11 |
| SARNet (Proposed) | 77.81 (+8.38%) | 0.111 | 2.01 | 231.69 (+10.09%) | 0.68 | 1.07 | 0.19 | 0.09 |
| 2. Time Domain Architecture: SSTA | ||||||||
| Baseline (RPB) | 74.18 | 0.550 | 273.70 | 168.56 | 0.55 | 2.93 | 1.11 | 0.14 |
| w/ Stem | 74.33 (+0.15%) | 0.550 | 34.26 | 180.75 (+6.74%) | 0.68 | 1.15 | 0.29 | 0.12 |
| w/ GSA | 75.21 (+1.03%) | 0.042 | 13.65 | 208.92 (+19.32%) | 1.05 | 42.98 | 1.03 | 0.48 |
| SSTA (Proposed) | 75.79 (+1.61%) | 0.042 | 1.75 | 229.58 (+26.58%) | 0.68 | 2.23 | 0.27 | 0.06 |
| 3. Fusion Architecture: RF-STARNet | ||||||||
| Baseline | 78.89 | 0.144 | 3.75 | 230.38 | 0.72 | 2.84 | 0.27 | 0.11 |
| w/ AGM | 80.41 (+1.52%) | 0.177 | 3.79 | 237.43 (+2.97%) | 0.70 | 4.06 | 0.27 | 0.11 |
| w/ Multi-Loss | 79.32 (+0.43%) | 0.154 | 3.76 | 232.30 (+0.83%) | 0.71 | 2.57 | 0.29 | 0.21 |
| RF-STARNet (Proposed) | 81.44 (+2.55%) | 0.188 | 3.80 | 242.99 (+5.19%) | 0.55 | 2.86 | 0.27 | 0.16 |
| Part II: Benchmark Comparison | ||||||||
| MobileNetV3 | 77.36 | 0.930 | 316.03 | 179.64 | 0.85 | 19.08 | 2.12 | 0.37 |
| ResNet18 | 76.11 | 11.180 | 9115.93 | 147.91 | 2.27 | - | 5.17 | 0.33 |
| ResNet-1D | 73.92 | 0.135 | 3.54 | 203.18 | 0.52 | 1.08 | 0.19 | 0.10 |
| S3R | 72.04 | 34.530 | 547.85 | 138.47 | 1.63 | - | 1.68 | 0.60 |
To investigate the internal working mechanism of RF-STARNet, we provide detailed visualizations of feature activation and error distribution.
We utilize CAM to visualize the regions of interest for different network branches.
-
SARNet: Displays distinct horizontal stripe patterns, indicating its focus on positional information of critical frequency blocks.
-
SSTA: Leverages global temporal correlations to produce precise masks covering all relevant frequency blocks (e.g., VTS and FCS signals), demonstrating superior noise resistance.
-
2D-CNN: Tends to perceive frequency blocks as a single holistic object, lacking signal-specific refinement compared to our specialized branches.
To validate the synergistic effect of the dual-branch fusion, we conducted a detailed error analysis using mosaic plots. Green pixels indicate samples where the fused model corrected the error of a single branch.
(a) Fusion vs. SARNet Branch |
(b) Fusion vs. SSTA Branch |
-
Fig. (a): RF-STARNet compensates for SARNet's deficiencies in lower SNR regions (dense green points), proving that temporal features are crucial when spectral features are corrupted by noise.
-
Fig. (b): While SSTA may perform suboptimally at high SNRs, the fused model maintains high accuracy by leveraging SARNet. In low-SNR conditions, RF-STARNet effectively exploits SSTA's robust global correlation features to ensure stability.
Conclusion: In this paper, we propose RF-STARNet, a multi-model lightweight architecture designed to address the trade-off between accuracy and computation overhead in UAV RF fingerprinting. By effectively decoupling the signal processing into stationary spectral and cyclostationary temporal domains, the proposed SARNet and SSTA modules extract sparse, complementary fingerprints with a lower overhead. Furthermore, the integration of an adaptive gating mechanism and a deep supervision strategy ensures robust feature fusion. Extensive experiments validate that SARNet, SSTA, and RF-STARNet achieve a superior balance between high recognition accuracy and low computational cost, exhibiting remarkable efficiency and robustness compared to comparison methods. These attributes make the proposed methods highly suitable for edge deployment. Future work will focus on the embedded implementation of this framework in complex, real-world electromagnetic environments.
If you find this work useful for your research, please consider citing:
@article{liu2026rf,
title={RF-STARNet: A Multi-Modal Lightweight Fusion Framework of UAV RF Fingerprinting Identification for Public Low-Altitude Safety},
author={Liu, Zhenkun and Mu, Junsheng and Hu, Panhe and Fan, Xing and Xu, Fenghua and Wang, Yin},
journal={IEEE Transactions on Vehicular Technology},
year={2026},
publisher={IEEE}
}
Zhenkun Liu |
Yunling Liao |
Zhiyv Dai |
This repository is under the Apache-2.0 license.