Few-Shot Specific Emitter Identification via Integrated Complex Variational Mode Decomposition and Spatial Attention Transfer

Specific emitter identification (SEI) utilizes passive hardware characteristics to authenticate transmitters, providing a robust physical-layer security solution. However, most deep-learning-based methods rely on extensive data or require prior information, which poses challenges in real-world scenarios with limited labeled data. We propose an integrated complex variational mode decomposition algorithm that decomposes and reconstructs complex-valued signals to approximate the original transmitted signals, thereby enabling more accurate feature extraction. We further utilize a temporal convolutional network to effectively model the sequential signal characteristics, and introduce a spatial attention mechanism to adaptively weight informative signal segments, significantly enhancing identification performance. Additionally, the branch network allows leveraging pre-trained weights from other data while reducing the need for auxiliary datasets. Ablation experiments on the simulated data demonstrate the effectiveness of each component of the model. An accuracy comparison on a public dataset reveals that our method achieves 96% accuracy using only 10 symbols without requiring any prior knowledge.

💡 Research Summary

The paper addresses the pressing challenge of Specific Emitter Identification (SEI) under few‑shot conditions, where only a handful of labeled signal samples are available. Traditional SEI approaches either rely on handcrafted time‑frequency transformations (e.g., STFT, Hilbert‑Huang, wavelet) or on deep‑learning models that demand large labeled datasets and often require prior knowledge such as modulation parameters or channel state information. Both strategies struggle in realistic IoT and cognitive‑radio scenarios where data collection is costly and device heterogeneity is high.

To overcome these limitations, the authors propose an integrated framework that combines four key components: (1) Integrated Complex Variational Mode Decomposition (ICVMD), (2) a Temporal Convolutional Network (TCN), (3) a Fully Convolutional Network (FCN) classifier equipped with a Spatial Attention Mechanism (SAM), and (4) a branch‑network for transfer learning from auxiliary datasets.

ICVMD extends the classic Variational Mode Decomposition (VMD) to the complex domain, allowing direct decomposition of I/Q samples without discarding phase information. By formulating a complex‑valued objective and enforcing analytic constraints, ICVMD separates the signal into intrinsic mode functions that collectively reconstruct an approximation of the original transmitted waveform. This reconstruction preserves subtle hardware‑induced distortions (the Radio Frequency Fingerprints, RFFs) that are essential for distinguishing emitters.

TCN replaces recurrent architectures with dilated causal convolutions, providing a large receptive field while maintaining computational efficiency. The TCN processes the sequence of mode‑wise features output by ICVMD, capturing long‑range temporal dependencies that arise from power‑amplifier memory effects, I/Q imbalance, and other non‑linearities.

FCN + SAM addresses the over‑fitting problem inherent to fully‑connected classifiers when training data are scarce. By using only convolutional layers, the classifier drastically reduces the number of trainable parameters and becomes invariant to the absolute position of informative signal segments. The Spatial Attention Mechanism learns a weighting map over the temporal (or frequency‑channel) dimension, emphasizing segments that contain strong RFF cues while suppressing noisy or redundant portions.

Branch‑Network Transfer enables the model to leverage pre‑trained weights from a related but not identical dataset. Only a subset of the network (the “branch”) is transferred, ensuring that generic low‑level features (e.g., PA non‑linearity patterns) are reused, while higher‑level layers are fine‑tuned on the few available target samples. This design mitigates negative transfer when domain gaps are large.

The authors conduct extensive experiments on both simulated data—generated by a detailed behavioral model of a power amplifier (Volterra series) and I/Q imbalance—and a public RFF dataset. Ablation studies demonstrate that each component contributes positively: ICVMD improves reconstruction fidelity, TCN boosts sequential modeling, SAM raises classification robustness, and the branch network reduces the need for auxiliary data. In a few‑shot test using only 10 symbols per emitter, the full system achieves 96 % accuracy, outperforming state‑of‑the‑art baselines (e.g., SRP‑CBL, MAML‑based SEI) which hover around 88–91 % under the same constraints. The method also remains resilient at low SNR (down to –5 dB).

In summary, the paper presents a novel, hardware‑aware signal decomposition technique combined with modern sequence modeling and attention‑driven classification, delivering a highly effective SEI solution that works with minimal labeled data and without extensive prior knowledge. The work opens avenues for real‑time deployment in resource‑constrained wireless networks and suggests future extensions to multi‑modal fingerprinting and broader modulation families.