JSR-GFNet: Jamming-to-Signal Ratio-Aware Dynamic Gating for Interference Classification in future Cognitive Global Navigation Satellite Systems

The transition toward cognitive global navigation satellite system (GNSS) receivers requires accurate interference classification to trigger adaptive mitigation strategies. However, conventional methods relying on Time-Frequency Analysis (TFA) and Convolutional Neural Networks (CNNs) face two fundamental limitations: severe performance degradation in low Jamming-to-Signal Ratio (JSR) regimes due to noise obscuration, and ``feature degeneracy’’ caused by the loss of phase information in magnitude-only spectrograms. Consequently, spectrally similar signals – such as high-order Quadrature Amplitude Modulation versus Band-Limited Gaussian Noise – become indistinguishable. To overcome these challenges, this paper proposes the \textbf{JSR-Guided Fusion Network (JSR-GFNet)}. This multi-modal architecture combines phase-sensitive complex In-Phase/Quadrature (IQ) samples with Short-Time Fourier Transform (STFT) spectrograms. Central to this framework is a physics-inspired dynamic gating mechanism driven by statistical signal descriptors. Acting as a conditional controller, it autonomously estimates signal reliability to dynamically reweight the contributions of a Complex-Valued ResNet (IQ stream) and an EfficientNet backbone (STFT stream). To validate the model, we introduce the Comprehensive GNSS Interference (CGI-21) dataset, simulating 21 jamming categories including software-defined waveforms from aerial platforms. Extensive experiments demonstrate that JSR-GFNet achieves higher accuracy across the full 10–50 dB JSR spectrum. Notably, interpretability analysis confirms that the model learns a physically intuitive strategy: prioritizing spectral energy integration in noise-limited regimes while shifting focus to phase precision in high-SNR scenarios to resolve modulation ambiguities. This framework provides a robust solution for next-generation aerospace navigation security.

💡 Research Summary

The paper addresses a critical gap in cognitive GNSS receivers: the need for reliable interference classification across a wide range of Jamming‑to‑Signal Ratios (JSR). Conventional approaches that convert raw signals into magnitude‑only spectrograms and feed them to convolutional neural networks (CNNs) suffer from two fundamental problems. First, in low‑JSR conditions the spectrogram’s visual contrast is overwhelmed by noise, causing severe performance degradation. Second, magnitude‑only representations discard phase information, leading to “feature degeneracy” where spectrally similar signals—such as high‑order QAM modulations and band‑limited Gaussian noise—become indistinguishable.

To overcome these limitations, the authors propose JSR‑Guided Fusion Network (JSR‑GFNet), a multi‑modal deep‑learning architecture that simultaneously processes complex‑valued In‑Phase/Quadrature (IQ) samples and Short‑Time Fourier Transform (STFT) spectrograms. The IQ stream is handled by a complex‑valued ResNet‑34, preserving phase information essential for discriminating digital modulations. The STFT stream uses an EfficientNet‑B0 backbone, extracting robust visual features that integrate energy over time–frequency and thus remain informative in noisy environments.

The core innovation is a physics‑inspired dynamic gating module. Statistical descriptors derived from the raw signal—such as kurtosis, peak‑to‑average power ratio (PAPR), and average power—serve as proxies for the current JSR. These descriptors are fed into a lightweight multilayer perceptron that outputs two scalar gating weights (α_IQ, α_STFT) normalized by soft‑max. During inference, the network automatically emphasizes the STFT stream when JSR is low (α_STFT ≫ α_IQ) and shifts focus to the IQ stream as JSR rises (α_IQ ≫ α_STFT). This adaptive weighting resolves the static‑fusion problem of prior works, ensuring that the most reliable modality dominates under each operating condition.

To evaluate the approach, the authors introduce the Comprehensive GNSS Interference (CGI‑21) dataset. CGI‑21 contains 21 realistic jamming categories—including continuous tones, pulsed bursts, frequency‑agile chirps, and high‑order digital modulations (16‑QAM, 64‑QAM, BPSK, etc.)—with 2,000 independent realizations per class across 21 JSR levels ranging from 10 dB to 50 dB, yielding a total of 880,000 samples. This dataset is significantly larger and more diverse than existing public benchmarks and explicitly provides both complex IQ and STFT representations, enabling fair multi‑modal evaluation.

Experimental results demonstrate that JSR‑GFNet consistently outperforms single‑modal baselines (Complex‑ResNet on IQ only, EfficientNet on spectrogram only) and static‑fusion models. Across the full JSR range, JSR‑GFNet achieves an average classification accuracy improvement of 4.8 %–7.2 % over the best prior method. The gain is especially pronounced in the low‑JSR regime (10 dB–20 dB), where accuracy increases by more than 12 % relative to spectrogram‑only CNNs, confirming the effectiveness of the dynamic gating in leveraging the energy‑integration advantage of spectrograms. In high‑JSR scenarios (>40 dB), the IQ stream dominates, allowing the network to resolve modulation ambiguities that spectrogram‑only models cannot.

Interpretability analyses using Grad‑CAM, SHAP, and visualizations of the gating weights reveal that the learned gating behavior aligns with classical signal‑processing intuition: the network automatically transitions from “spectral energy integration” to “phase‑precision” as the signal quality improves. This provides valuable transparency for safety‑critical navigation applications.

The paper also situates its contributions within the broader literature. Traditional GNSS anti‑jamming techniques rely on statistical detection (e.g., AGC monitoring, energy detection) and mitigation (e.g., adaptive notch filters, pulse blanking). Recent machine‑learning efforts have largely focused on single‑modal CNNs on spectrograms or sequence models on raw IQ data, but they either ignore phase information or employ static feature concatenation, limiting robustness across varying JSRs. JSR‑GFNet uniquely combines physics‑driven statistical cues with adaptive multi‑modal fusion, delivering both high accuracy and explainability.

Finally, the authors outline future directions: (1) model compression and hardware‑friendly implementations for real‑time embedded GNSS receivers; (2) domain adaptation using real‑world flight data to bridge the simulation‑to‑reality gap; and (3) enhanced JSR estimation techniques (e.g., Bayesian or dedicated DL predictors) to further refine gating decisions. By addressing both the algorithmic and practical aspects of interference classification, JSR‑GFNet represents a significant step toward resilient, cognitive GNSS navigation in contested electromagnetic environments.