Reliability-Guided Spectrum Generative Augmentation for Few-Shot Emitter Identification in Mobile Networks

Technical Summary: Reliability-Guided Spectrum Generative Augmentation for Few-Shot Emitter Identification

Problem Statement
In next-generation mobile networks, the electromagnetic spectrum functions as an interactive environment where communication, sensing, and autonomous control are coupled. Specific Emitter Identification (SEI) is critical for securing this environment by distinguishing transmitters via hardware-dependent radio-frequency fingerprints (RFFs). However, in dynamic scenarios such as low-altitude mobile networks or self-organizing infrastructures, newly accessed or intermittently visible emitters often provide only a few labeled observations. Conventional deep learning models, which assume abundant labeled data, struggle in these few-shot conditions, often fitting incidental noise or channel distortions rather than stable RFF characteristics. Furthermore, while generative augmentation can expand limited datasets, treating synthetic samples as fully trusted observations risks biasing decision boundaries and propagating incorrect identity information into downstream spectrum decisions.

Methodology: RSGA-SEI
The paper proposes RSGA-SEI (Reliability-Guided Spectrum Generative Augmentation), a framework designed for "embodied spectrum intelligence." The approach operates in a closed-loop workflow comprising three main stages:

1. CSI Representation: Baseband observations are transformed into Contour Stella Image (CSI) samples. This maps in-phase (I) and quadrature (Q) sampling distributions into 2D density patterns, preserving amplitude, phase, and local density information in a visual format suitable for deep learning.
2. SAGA (Spectrum-Aware Generative Augmentation): This module synthesizes auxiliary CSI samples to expand the support set. It employs:
   • SCPE (Spectrum-Context Prior Encoder): Extracts both local contour/texture cues and global spatial context from CSI images.
   • NVDR (Neighbor-View Diversity Regularization): Prevents the generator from producing redundant images by enforcing diversity in neighboring latent views while maintaining semantic alignment.
   • DPCD (Dual-Prior Consistency Discrimination): Uses local and context-aware discriminators to ensure generated samples are physically consistent with real data, minimizing disagreement between prior views.
3. SE-RWC (Spectrum-Embedded Reliability-Weighted Classifier): Unlike methods that optimize prototype geometry, SE-RWC uses measured support samples as the sole "trust anchor."
   • It computes class prototypes and dispersion solely from real data.
   • Generated samples are assigned soft reliability weights based on their feature distance to the real prototypes. Samples with high consistency receive higher weights, while inconsistent synthetic data is down-weighted.
   • The classifier is trained using a weighted cross-entropy loss, ensuring that unreliable synthetic evidence does not distort the decision boundary.
   • Crucially, generation and reliability scoring occur only during training. The online inference path remains compact, utilizing only the trained lightweight classifier.

Key Contributions

• Formulation: The authors frame few-shot SEI as a spectrum-embedded recognition problem, linking identification with perception, decision, and action feedback loops.
• Framework Design: The introduction of RSGA-SEI, which converts baseband data to CSI, expands the support set via spectrum-aware synthesis, and calibrates the contribution of synthetic data through reliability weighting.
• Reliability Calibration: The design of SE-RWC, which avoids heavy prototype-geometry optimization. Instead, it uses real support prototypes to assign soft weights to generated samples, keeping the deployment path simple.
• Comprehensive Evaluation: The study validates the method through 5-shot comparisons, ablation studies, multi-shot generalization, SNR robustness analysis, and complexity indicators.

Experimental Results
Evaluated on the NJUPT-RFF dataset (7 emitters, 5-shot training, 50-sample query):

• Accuracy: RSGA-SEI achieved 87.14% accuracy in the 5-shot setting. This represents a 13.71 percentage point improvement over the non-augmented CSI baseline (73.43%) and a 3.43 percentage point gain over using SAGA alone without reliability weighting (83.71%).
• Ablation: The study confirmed that the primary bottleneck is sample scarcity (large gain from CSI to CSI+SAGA) and that reliability should be a graded signal rather than a binary filter (progressive gains from hard filtering to soft weighting to full SE-RWC).
• Robustness: The method maintained superior performance across varying Signal-to-Noise Ratios (SNR), particularly in low-to-medium SNR ranges (e.g., +3.14% over CSI-DA at 5 dB), demonstrating that reliability weighting effectively suppresses harmful synthetic samples in noisy conditions.
• Stability: As the support set size increased (5 to 30 shots), Cohen's Kappa improved from 0.850 to 0.940, and the coefficient of variation decreased, indicating stable prototype reliability estimates.
• Complexity: While training is computationally heavier due to generation and calibration, the online inference path remains compact (ResNet-18 backbone). Inference latency increased only marginally from 3.6 ms to 3.7 ms, with parameter counts remaining at 11.18M.

Significance and Claims
The paper claims that RSGA-SEI addresses a critical gap in embodied spectrum intelligence: the need for robust few-shot learning that does not compromise the integrity of the decision loop with unreliable synthetic data. By decoupling the generation/calibration phase from the online inference phase, the framework supports adaptive mobile systems where recognition outputs directly influence spectrum access and resource coordination. The authors emphasize that their approach connects spectrum perception, sample expansion, and emitter recognition in a compact closed-loop workflow, offering a practical solution for scenarios where labeled data is scarce but real-time, reliable identification is essential. The method is presented as a step toward secure, self-organizing wireless infrastructures, though the authors modestly note limitations regarding extremely small support sets or missing physical modes in the generator.