JaGNet: Detection, Identification and Position Error Correction of GNSS Jamming with Deep Temporal Graphs

Ivana Kesić; Aljaž Blatnik; Carolina Fortuna; Blaž Bertalanič

cs.LGSep 17, 2025

JaGNet: Detection, Identification and Position Error Correction of GNSS Jamming with Deep Temporal Graphs

Authors: Ivana Kesić, Aljaž Blatnik, Carolina Fortuna, Blaž Bertalanič

Organizations: Department of Communication Systems, Jožef Stefan Institute, Slovenia · Faculty of Electrical Engineering, University of Ljubljana, Slovenia

Abstract

Global Navigation Satellite Systems (GNSS) underpin location-dependent mobile computing, such as railway signaling, public transport and fleet tracking, yet face growing disruption from low-cost jammers. Under moderate jamming, low-cost receivers keep reporting a valid fix, shifted by up to several meters and often without an interference flag. Their built-in defenses run in closed firmware, and existing data-driven position error correction methods are designed for multipath, require raw pseudoranges or operate on single epochs. We propose JaGNet, a single-layer heterogeneous graph convolutional LSTM over star graphs of the receiver and its satellites, which, from the standard receiver output alone, detects the jamming, identifies its type and power, and corrects the reported position only when jamming is detected. We evaluate it on new hardware-in-the-loop recordings of a commercial multi-constellation receiver under three jammer types at six powers, at six static locations and on eleven railway and bus routes. JaGNet detects 99.8 to 99.9% of the jammed 10-s windows with at most 1.0% false positives, against 64.7 to 66.2% for the receiver's own monitor, and identifies the jammer type in 95.2 to 96.6% of them. Trained in the laboratory only, it detects 97.7% of the Jammertest 2024 field windows with a C/N0 drop of 10 dB or more. It reduces the mean absolute position error by 67.5% at held-out locations and by 64.0% on held-out route repetitions, where its error is 20.2 cm below that of the strongest baseline, whereas at the static locations the attention-based baselines are within 0.4 cm of it.

Figures & tables

Fig. 1 : JaGNet as a software layer between a low-cost GNSS receiver and the application. Under jamming (1), the receiver keeps reporting a valid but shifted fix (2). From the standard output alone, JaGNet identifies the jamming and corrects the position (3) for the application (4).

Fig. 2 : Three distinct jamming profiles, represented in different colors, overlaid on the L1-band GNSS spectrum (gray shaded area), illustrating the significant power disparity between the interference and the nominal signal.

Fig. 3 : Transformation of the physical GNSS constellation into a heterogeneous star-graph snapshot. The receiver is modeled as the central node, while visible satellites $S_{t}$ are represented as leaf nodes.

Panel of Figure 4 — Fig. 4 : Comparison of constellation temporal graph sequences over a sliding window from $t-k$ to $t$ .

Fig. 5 : Architecture of JaGNet . A single HeteroGCLSTM layer processes the last $k$ graph snapshots, and the 64-dimensional state of the receiver feeds the classification head ( $\hat{c}_{t}$ ) to identify jamming and the correction head ( $\hat{Y}_{t}$ ) to do PEC. Absent (hollow) satellites keep their frozen states.

Fig. 6 : Methodology: laboratory and open field data (1), graph construction (2), training and held-out evaluation (3), and comparison of JaGNet with the baselines (4).

Fig. 7 : Positional deviation of a GNSS receiver across 50 repeated measurements under

3\times

CW jamming from 100 to 200 s (shaded), adapted from [ 21 ] .

	Static					Kinematic
	Jammed windows (%) $\uparrow$					Jammed windows (%) $\uparrow$
Method	Det.	FM	Type	5 dB	Exact	Det.	FM	Type	5 dB	Exact
JaGNet , by jammer power
$-45$ dBm	100.0	100.0	98.8	76.8	64.0	100.0	100.0	99.5	71.0	56.6
$-50$ dBm	100.0	100.0	96.6	83.3	55.1	100.0	100.0	97.2	74.5	48.2
$-55$ dBm	100.0	100.0	96.0	85.7	61.6	100.0	100.0	92.9	85.0	47.2

TABLE II : Identification in the laboratory (%, pooled over ten seeds). FM: detected under FM jamming. 5 dB: type correct, power within 5 dB. Exact: type and power correct. Bold, over all powers except FM: highest of the column and the values not significantly below it (paired 95% bootstrap CI, Section V-E ).

\uparrow

\downarrow

: higher/lower is better.

	Jammertest 2024				Laura Bassi cruise
		Detected $\uparrow$				Detected $\uparrow$
Method	FP $\downarrow$	3–6	6–10	$\geq$ 10	FP $\downarrow$	3–6	$>$ 6
$n$	64	31	41	137	4,539	125	28
JaGNet	0.0	23.2	65.4	97.7	1.5	25.0	100.0
GCNN	0.0	3.2	42.0	100.0	0.5	21.4	100.0
TSMixer	60.6	62.6	66.8	94.2	40.5	39.2	98.6

TABLE III : Detection over the air by models trained in the laboratory only (%, mean over 5 seeds). FP: false-positive rate on the clean Jammertest windows and cruise minutes. Detected: detection rate by the drop of the mean GPS C/N 0 (dB). Gray: more than 5% FP, so the detection rates are inflated and not compared. Bold: highest detection rate outside gray. –: not available.

Jamming	Method (MAE $\downarrow$ )	$-$ 45 dBm		$-$ 50 dBm		$-$ 55 dBm		$-$ 60 dBm		$-$ 65 dBm		$-$ 70 dBm
CW	Proposed - JaGNet	118.3 $\pm$	34.0	72.1 $\pm$	15.3	45.6 $\pm$	7.1	30.8 $\pm$	7.0	24.8 $\pm$	4.4	37.3 $\pm$	3.8
	SotA - GCNN [ 30 ]	119.3 $\pm$	31.7	94.4 $\pm$	13.5	50.3 $\pm$	7.2	32.5 $\pm$	6.3	33.0 $\pm$	8.6	41.5 $\pm$	4.4
	Baseline - TSMixer	121.9 $\pm$	29.6	87.6 $\pm$	15.9	48.9 $\pm$	5.0	33.4 $\pm$	7.2	28.9 $\pm$	4.1	37.1 $\pm$	3.6
	Baseline - Set Transformer	118.8 $\pm$	35.1	74.8 $\pm$	14.6	44.8 $\pm$	7.1	31.6 $\pm$	7.0	25.6 $\pm$	4.8	38.9 $\pm$	3.3
	Baseline - Set Transformer + LSTM	119.0 $\pm$	36.0	69.4 $\pm$	15.0	43.7 $\pm$	7.8	30.9 $\pm$	7.0	24.7 $\pm$	5.1	37.3 $\pm$	4.3
	Unmitigated Error (cm)	215.3 $\pm$	42.3	229.8 $\pm$	15.9	124.9 $\pm$	6.0	114.3 $\pm$	10.8	112.2 $\pm$	5.6	122.9 $\pm$	7.3

TABLE IV : Euclidean MAE (cm) at the held-out locations per jammer type and power: mean over the six locations and ten seeds

\pm

standard deviation over the locations. Bold: lowest MAE of the setting and the MAEs not significantly above it (paired 95% bootstrap CI, Section V-E ).

Route (MAE $\downarrow$ )	Uncorr.	JaGNet	GCNN	TSMixer	Set Tr.	Set Tr. + LSTM
FE–UKC (bus)	768.9	238.9	712.4	374.6	343.2	293.7
Medvode–Topol (bus)	804.5	257.1	729.4	340.5	347.7	312.6
T01 Ljubljana Črnuče–Domžale	245.5	76.9	127.1	103.6	95.3	87.7
T02 Domžale–Ljubljana Črnuče	220.2	85.4	120.0	105.9	95.2	103.8
T03 Ptuj–Moškanjci	281.9	90.9	105.9	98.1	95.8	94.9
T04 Ljubljana–Ljubljana Zalog	216.6	77.2	98.7	87.6	85.6	81.8

TABLE V : Euclidean MAE (cm) on the held-out repetition of each route, for models trained on repetitions 1–3 of the same route. Bold: lowest MAE of the route, or of the mean over the routes, and the MAEs not significantly above it (paired 95% bootstrap CI, Section V-E ).

	Static			Kinematic
Method	MAE $\downarrow$	$R$ $\uparrow$	$\Delta$ MAE [95% CI]	MAE $\downarrow$	$R$ $\uparrow$	$\Delta$ MAE [95% CI]
Uncorrected	174.9	–	–	334.7	–	–
Models
JaGNet	56.9	67.5	–	120.3	64.0	–
GCNN [ 30 ]	59.8	65.8	2.6 [2.0, 3.4]	229.8	31.3	108.6 [83.6, 134.5]
TSMixer	59.6	65.9	2.5 [1.7, 3.2]	155.0	53.7	33.4 [27.2, 39.9]

TABLE VI : Mean Euclidean MAE (cm) over the held-out locations or routes, error reduction

R

(%), and paired MAE difference to JaGNet over the test recordings (

\Delta

MAE, cm, with 95% bootstrap CI, positive when JaGNet is lower), which can differ from the difference of the MAE columns by up to 1.3 cm. Oracles use the true jamming setting.

	FE–UKC (bus) , $R$ $\uparrow$		T07 (railway) , $R$ $\uparrow$
$k$	JaGNet	Set Tr. + LSTM	JaGNet	Set Tr. + LSTM
0	3.3	1.8	33.5	27.5
4	38.8	18.0	14.9	12.6
10	56.2	31.7	47.3	37.1
19	65.1	53.1	48.8	43.8
38	68.1	58.4	60.9	52.1

TABLE VII : Calibration along a new route: error reduction

R

(%) on the held-out repetition of a model trained from scratch on

k

recordings of the route (

k=0

: model of the ten other routes).

	Static				Kinematic
	$R$ (%) $\uparrow$				$R$ (%) $\uparrow$
Left out	With	W/o	Ratio $\uparrow$	Det. $\uparrow$	With	W/o	Ratio $\uparrow$	Det. $\uparrow$
CW	64.2	37.8	0.59	80.3	59.4	29.0	0.49	77.7
3 $\times$ CW	64.5	63.6	0.99	100.0	66.0	61.2	0.93	100.0
FM	72.4	64.9	0.90	100.0	69.0	36.9	0.53	100.0
$-45$ dBm, all types	46.9	43.8	0.93	100.0	49.8	41.8	0.84	100.0

TABLE VIII : Jammer type or power left out of training: error reduction

R

(%) of JaGNet on the left-out setting with and without it in training, their ratio, and the detection rate (%) of the model trained without it.

		MAE $\downarrow$		Exact setting $\uparrow$
Model	Sat.	Static	Kinematic	Static	Kinematic
JaGNet	GPS	56.9	120.3	63.5	60.1
JaGNet	all	54.5	112.1	95.4	93.2
Set Tr. + LSTM	GPS	56.5	140.6	65.7	61.0
Set Tr. + LSTM	all	54.4	137.4	96.7	93.5

TABLE IX : Two-head models with the satellites of GPS only (main results) or of all constellations (GPS, Galileo, BeiDou) as nodes, on the same windows: MAE (cm) and exact jamming setting identified (%).

	Static				Kinematic, FE–UKC
Variant	$R$ $\uparrow$	$\Delta$ MAE [95% CI]	Type $\uparrow$	Exact $\uparrow$	$R$ $\uparrow$	$\Delta$ MAE [95% CI]	Type $\uparrow$	Exact $\uparrow$
Proposed ( $W=10$ , hidden size 64)	68.2	–	97.5	64.6	69.1	–	97.0	57.9
Inputs
w/o SNR	62.3	10.4 [8.8, 12.0]	83.7	27.3	61.4	62.7 [26.7, 104.6]	84.0	28.3
w/o elevation	68.1	0.3 [0.1, 0.5]	97.2	64.3	69.1	$-$ 0.0 [ $-$ 3.7, 3.4]	97.5	57.7
w/o azimuth	68.2	0.1 [ $-$ 0.1, 0.2]	97.1	64.0	68.8	2.5 [ $-$ 1.4, 6.4]	95.5	55.4

TABLE X : Ablation of JaGNet at rest and on the route FE–UKC: error reduction

R

(%), paired MAE difference to the proposed model (

\Delta

MAE, cm, mean and 95% bootstrap CI, positive when the variant is worse), and type and exact-setting accuracy (%).

			Latency (ms) $\downarrow$
Model	Param. $\downarrow$	MFLOP $\downarrow$	Rest	Motion	p99
JaGNet	69,205	5.8	37.3	37.7	38.9
GCNN [ 30 ]	5,111	0.1	1.2	1.1	1.2
TSMixer	47,050	0.7	1.4	1.4	1.4
Set Transformer	90,453	13.3	6.1	6.2	7.6
Set Tr. + LSTM	89,941	9.2	4.2	4.3	4.8

TABLE XI : Inference cost on a Raspberry Pi 5 per 10-s window: mean latency at rest and in motion, and 99th percentile in motion.

Spatiotemporal GNNs

Spatiotemporal graph neural networks model how signals and relationships evolve across connected locations, agents, or physical systems.

See all 147 papers

Sep 28, 2026

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

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.

Zhihan Zeng, Kaihe Wang, José A. López-Salcedo +2

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

Multi-Label Classification Compositional Representation Learning Compositional Generalization

Ask PDF

May 14, 2026

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

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.

Thorben Wegner, Lucas Heublein, Tobias Feigl +3

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

Variational Autoencoders Energy-Efficient ML Edge ML

Ask PDF

May 12, 2026

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

Global navigation satellite system (GNSS) interference poses a serious threat to reliable positioning, especially in indoor and multipath-rich environments where source localization is highly challenging. In this paper, we formulate GNSS interference localization as an active sensing problem and propose a reinforcement learning (RL) framework in which an agent sequentially explores the environment to infer the position of an emitter source from radio frequency (RF) observations acquired with a 2x2 patch antenna. The localization task is modeled as a partially observable decision process, since single-snapshot measurements are often ambiguous under multipath propagation and changing channel conditions. To address this, the proposed framework combines high-dimensional RF sensing with deep RL and recurrent policy learning. We investigate both value-based and policy-based approaches, namely Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO), and study their behavior under domain shift. The approach is evaluated on a simulated dataset generated with the Sionna ray-tracing module, which provides realistic propagation effects and diverse environment configurations. Experimental results show that the proposed method achieves a localization success rate of 80.1, demonstrating the potential of RL for adaptive GNSS interference localization. Overall, the results highlight simulation-assisted training as a promising direction for robust interference localization in challenging propagation environments.

M. Shamail J. Khan, Nisha L. Raichur, Lucas Heublein +5

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

Active Sensing Wireless Localization

Ask PDF

Abstract

Spatiotemporal GNNs

Explore similar work

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

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

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