SPOC-Net: Single-Primitive Online Composition Network for GNSS Jamming Set Recognition

Zhihan Zeng; Kaihe Wang; José A. López-Salcedo; Gonzalo Seco-Granados; Zhongpei Zhang

cs.CVSep 28, 2026

SPOC-Net: Single-Primitive Online Composition Network for GNSS Jamming Set Recognition

Authors: Zhihan Zeng, Kaihe Wang, José A. López-Salcedo, Gonzalo Seco-Granados, Zhongpei Zhang

Organizations: National Key Laboratory of Wireless Communications, University of Electronic Science and Technology of China (UESTC), Chengdu, China · Shenzhen Institute for Advanced Study, UESTC, Shenzhen 518110, China · Universitat Aut`onoma de Barcelona, 08193 Barcelona, Spain

Abstract

Reliable positioning, navigation, and timing support intelligent transportation, autonomous systems, and space-air-ground integrated networks. However, global navigation satellite system (GNSS) jamming recognizers that treat each mixture as a separate class are difficult to extend to new combinations. Therefore, this paper proposes SPOC-Net, which decomposes the recognition problem into identifying a set of basic jamming components. Multi-resolution time-frequency features and learned component queries provide evidence for each component type. A high-resolution branch estimates the number of active types, and a structured decoder combines this estimate with component evidence to select a valid set. For training, measured single-component records are the only physical samples used in gradient optimization. Their associated clean in-phase and quadrature (IQ) sequences are combined on demand during training to produce labeled mixtures with different relative powers and jamming-to-noise ratios. Separate measured mixtures from ten training-listed compositions support model selection and decoder calibration; six other compositions are reserved for final testing. Evaluation on 14,220 independently generated, conductively combined, and recorded radio frequency mixtures yields 80.69% exact-set accuracy and a 92.84% micro-averaged F1 score. On combinations excluded from model development, SPOC-Net achieves 80.89% exact-set accuracy, exceeding the strongest comparison method by 18.77 percentage points under the reported protocols.

Figures & tables

Fig. 1: Conceptual GNSS jamming scenario motivating the received-signal model.

Fig. 2: Representative time-frequency representations of the five primitive jamming signals and their two-component and three-component compositions.

Sample source	Gradient optimization	Model selection	Decoder calibration	Final test
Measured singleton training records	Yes	No	No	No
Online mixed auxiliaries from $\mathcal{A}_{\mathrm{list}}$	Yes	No	No	No
Measured development records from $\mathcal{A}_{\mathrm{list}}$	No	Yes	Yes	No
Measured final-test records from $\mathcal{A}_{\mathrm{list}}$	No	No	No	Yes
Measured final-test records from $\mathcal{A}_{\mathrm{hold}}$	No	No	No	Yes

TABLE I: Roles of Physical, Auxiliary, Development, and Test Samples

Fig. 3: Architecture of SPOC-Net. A shared multi-resolution spectro-temporal encoder feeds a primitive-evidence branch and a detached high-resolution cardinality branch. The singleton-versus-mixture output provides auxiliary training supervision. Final inference is restricted to the 16 valid two-component and three-component masks and uses the conditional two-versus-three cardinality probability.

Setting	Value
Network input	$224\times 224$ grayscale STFT only
Coordinate augmentation	Time and frequency maps
Stem kernels and stride	$3,5,11$ and $4$
Stage depths	$(2,2,6,2)$
Stage channels	$(64,128,256,384)$
Decoder dimension $d$	$384$

TABLE II: Architecture and Decoder Configuration

Category	Setting	Category	Setting
Optimization	120 epochs; batch size 32	Optimizer	AdamW
Learning rate	$2.5\times 10^{-4}$ to $10^{-6}$	Weight decay	$7\times 10^{-4}$
Gradient handling	Norm clipping at $1.0$	EMA decay	$0.999$
Physical gradient data	Measured singleton records only	Auxiliary source	Clean-IQ singleton bank only
Training mixed masks	$\mathcal{A}_{\mathrm{list}}$ only	Training candidate universe	$\mathcal{A}_{1}\cup\mathcal{A}_{\mathrm{list}}$
Development compositions	$\mathcal{A}_{\mathrm{list}}$ only	Held-out development records	None

TABLE III: Training, Auxiliary Composition, Loss, Protocol, and Calibration Settings

Fig. 4: Conducted RF acquisition platform. Panel (a) shows the hardware implementation with two USRP X410 units, a GNSS antenna, and a four-input RF combiner. Panel (b) shows the signal flow: the live GNSS antenna signal and up to three jammer transmit channels are combined in the RF domain, recorded by the USRP receive channel, and passed to the IQ-processing pipeline.

Partition	Classes	$N_{\mathrm{rec}}$	Component sets
Training-listed final test	10	8,887	STJ+LFMJ, STJ+PBNJ, MTJ+LFMJ, MTJ+PTJ, LFMJ+PTJ, PTJ+PBNJ, STJ+LFMJ+PBNJ, STJ+PTJ+PBNJ, MTJ+LFMJ+PTJ, MTJ+LFMJ+PBNJ
Held-out final test	6	5,333	STJ+PTJ, MTJ+PBNJ, LFMJ+PBNJ, STJ+LFMJ+PTJ, MTJ+PTJ+PBNJ, LFMJ+PTJ+PBNJ
Full mixed final test	16	14,220	Union of the training-listed and held-out final-test partitions

TABLE IV: Composition-Level Final-Test Partitions

Category	Setting	Category	Setting
Prediction task	Five-label primitive-set prediction	Output and loss	Five logits, sigmoid, and ASL
Gradient data	Measured singletons and measured $\mathcal{A}_{\mathrm{list}}$ mixtures	Online composition	None
Held-out development data	None	Final candidate universe	All 16 valid two-component and three-component masks
Primary image input	$224\times 224$ grayscale STFT	JNR side information	None for every model
Initialization	Random, without ImageNet pretraining	Optimization	AdamW, 120 epochs, batch size 32
Augmentation	Shift, flip, jitter, noise, erase, and stripe mask	Sampling	Composition-balanced with higher three-component and low-JNR weight

TABLE V: Training and Inference Protocol for the Component-Set Comparison Models

Test data	$N_{\mathrm{rec}}$	Exact-set (%)	Micro-precision (%)	Micro-recall (%)	Micro-F1 (%)	Cardinality accuracy (%)	Hamming loss (%)
Full mixed set	14,220	80.69	95.31	90.49	92.84	87.08	6.81

TABLE VI: Overall SPOC-Net Results on the Full Measured Mixed Set

JNR (dB)	$N_{\mathrm{rec}}$	Exact-set (%)	Micro-F1 (%)	Two-component exact-set (%)	Three-component exact-set (%)
$-20$	1,778	16.59	58.85	29.28	0.13
$-15$	1,773	51.55	85.56	82.65	11.60
$-10$	1,777	81.99	95.84	96.10	63.79
$-5$	1,776	96.11	99.18	99.40	91.93
0	1,778	99.49	99.90	99.70	99.23
5	1,777	99.72	99.94	99.90	99.49

TABLE VII: SPOC-Net Performance Versus JNR on the Full Mixed Set

Model	Training-listed 10 classes (%)	Held-out 6 classes (%)	Full 16 classes (%)
AlexNet-ML	75.48	34.97	60.29
ResNet18-ML	79.00	62.12	72.67
Feature-ResNet18-ML	80.45	59.48	72.59
ACSNet-ML	78.87	43.95	65.77
MSFF-KAN-ML	82.00	50.85	70.32
TSFANet-ML	81.69	41.83	66.74

TABLE VIII: Exact-Set Accuracy of Five-Label Component-Set Models Across Composition Partitions

Fig. 5: Exact-set accuracy versus JNR on the 10 training-listed mixed compositions. The panels report overall, two-component, and three-component results for 8,887 measured records. The figure legend omits the “-ML” suffix for comparison models.

JNR (dB)	SPOC-Net (%)	MSFF-KAN-ML (%)	TSFANet-ML (%)
$-20$	0.2	2.9	20.0
$-15$	10.9	40.4	43.3
$-10$	61.0	80.4	76.6
$-5$	92.1	96.4	93.9

TABLE IX: Three-Component Accuracy on Training-Listed Compositions at Low JNR

Fig. 6: Row-normalized composition confusion matrices on the full measured mixed set. The upper row contains AlexNet-ML, ResNet18-ML, and Feature-ResNet18-ML. The middle row contains ACSNet-ML, MSFF-KAN-ML, and TSFANet-ML. The lower panel contains SPOC-Net. All 16 two-component and three-component classes and all tested JNR values are included. The artwork omits the “-ML” suffix.

Composition	$N_{\mathrm{rec}}$	Exact-set (%)	Micro-F1 (%)
$\mathrm{STJ}+\mathrm{PTJ}$	889	87.74	93.51
$\mathrm{MTJ}+\mathrm{PBNJ}$	890	98.76	99.47
$\mathrm{LFMJ}+\mathrm{PBNJ}$	889	84.81	92.77
$\mathrm{STJ}+\mathrm{LFMJ}+\mathrm{PTJ}$	889	76.49	89.84
$\mathrm{MTJ}+\mathrm{PTJ}+\mathrm{PBNJ}$	888	67.91	94.13
$\mathrm{LFMJ}+\mathrm{PTJ}+\mathrm{PBNJ}$	888	69.59	90.24

TABLE X: SPOC-Net Results on the Six Held-Out Compositions

Primitive	Precision (%)	Recall (%)	F1 (%)
STJ	98.99	90.29	94.44
MTJ	85.34	95.72	90.23
LFMJ	99.59	87.62	93.22
PTJ	99.86	90.50	94.95
PBNJ	92.57	90.00	91.26

TABLE XI: Primitive-Wise Performance on the Full Mixed Set

True cardinality	Predicted 2 (%)	Predicted 3 (%)
2	99.48	0.52
3	28.86	71.14

TABLE XII: Row-Normalized Cardinality Confusion

True set	Predicted set	Count
$\{\mathrm{MTJ},\mathrm{LFMJ},\mathrm{PBNJ}\}$	$\{\mathrm{MTJ},\mathrm{PBNJ}\}$	130
$\{\mathrm{MTJ},\mathrm{PTJ},\mathrm{PBNJ}\}$	$\{\mathrm{MTJ},\mathrm{PTJ}\}$	128
$\{\mathrm{STJ},\mathrm{PTJ},\mathrm{PBNJ}\}$	$\{\mathrm{STJ},\mathrm{PTJ}\}$	119
$\{\mathrm{MTJ},\mathrm{LFMJ},\mathrm{PBNJ}\}$	$\{\mathrm{MTJ},\mathrm{LFMJ}\}$	112
$\{\mathrm{STJ},\mathrm{LFMJ},\mathrm{PBNJ}\}$	$\{\mathrm{STJ},\mathrm{LFMJ}\}$	112

TABLE XIII: Most Frequent Set Errors

Explore similar work

GenAI for Energy-Efficient and Interference-Aware Compressed Sensing of GNSS Signals on a Google Edge TPU
May 14, 2026Thorben Wegner, Lucas Heublein, Tobias Feigl +3Intentional Jamming Global Navigation Satellite System
Fraunhofer Institute for Integrated Circuits IIS, 90411 Nürnberg, Germany · Friedrich-Alexander-Universität Erlangen-Nürnberg, 91058 Erlangen, Germany
JointHRRP-Net: A Statistically Constrained Decoupling Network for Joint Target and Jamming Recognition in Composite Jamming
May 19, 2026Yunfei Zhao, Mei Liu, Shuowei Liu +2Intentional Jamming Joint Search
College of Electronic Science and Technology, National University of Defense Technology, Changsha 410073, China
Active Sensing with Meta-Reinforcement Learning for Emitter Localization from RF Observations
May 12, 2026M. Shamail J. Khan, Nisha L. Raichur, Lucas Heublein +5Global Navigation Satellite System Localization
Fraunhofer Institute for Integrated Circuits IIS, 90411 Nürnberg, Germany · Center for Artificial Intelligence, Technical University of Applied Sciences Würzburg-Schweinfurt, Germany · Machine Learning and Positioning Systems Lab, University of Technology Nürnberg (UTN), 90461 Nürnberg, Germany +1

May 14, 2026, cs.LGJ/K move · Enter open · S save

GenAI for Energy-Efficient and Interference-Aware Compressed Sensing of GNSS Signals on a Google Edge TPU

Thorben Wegner, Lucas Heublein, Tobias Feigl, Felix Ott+2

Fraunhofer Institute for Integrated Circuits IIS, 90411 Nürnberg, Germany · Friedrich-Alexander-Universität Erlangen-Nürnberg, 91058 Erlangen, Germany

Intentional Jamming Global Navigation Satellite System Signal Reconstruction Interference Energy-Efficient Generative Artificial Intelligence

Open paper AskPDF arXiv

Traditional methods for classifying global navigation satellite system (GNSS) jamming signals typically involve post-processing raw or spectral data streams, requiring complex and costly data transmission to cloud-based interference classification systems. In contrast, our proposed approach efficiently compresses GNSS data streams directly at the hardware receiver while simultaneously classifying jamming and spoofing attacks in real time. Given the growing prevalence of GNSS jamming, there is a critical need for real-time solutions suitable for power-constrained environments. This paper introduces a novel method for compressing and classifying GNSS jamming threats using generative artificial intelligence (GenAI), specifically variational autoencoders (VAEs), deployed on Google Edge tensor processing units (TPUs). The study evaluates various autoencoder (AE) architectures to compress and reconstruct GNSS signals, focusing on preserving interference characteristics while minimizing data size near the receiver hardware. The pipeline adapts large-scale AE models for Google Edge TPUs through 8-bit quantization to ensure energy-efficient deployment. Tests on raw in-phase and quadrature-phase (IQ) data, Fast Fourier Transform (FFT) data, and handcrafted features show the system achieves significant compression (>42x) and accurate classification of approximately 72 interference types on reconstructed signals (F2-score 0.915), closely matching the original signals (F2-score 0.923). The hardware-centric GenAI approach also substantially reduces jammer signal transmission costs, offering a practical solution for interference mitigation. Ablation studies on conditional and factorized VAEs (i.e., FactorVAE) explore latent feature disentanglement for data generation, enhancing model interpretability and fostering trust in machine learning (ML) solutions for sensitive interference applications.

Abstract

Explore similar work

GenAI for Energy-Efficient and Interference-Aware Compressed Sensing of GNSS Signals on a Google Edge TPU

JointHRRP-Net: A Statistically Constrained Decoupling Network for Joint Target and Jamming Recognition in Composite Jamming

Active Sensing with Meta-Reinforcement Learning for Emitter Localization from RF Observations