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

arXiv:2509.14000 · cs.LG · Submitted 2025-09-17 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "JaGNet: Detection, Identification and Position Error Correction of GNSS Jamming with Deep Temporal Graphs".

Jane: The gist: JaGuard proposes receiver-centric deep temporal graph networks to correct position errors caused by GNSS jamming by reformulating jamming mitigation as a dynamic temporal graph regression problem.

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we’re diving into "JaGNet: Detection, Identification and Position Error Correction of GNSS Jamming with Deep Temporal Graphs." The title tells you immediately that this work moves beyond simple error correction to a whole new way of modeling the problem.

Jane: It focuses on jamming mitigation using deep temporal graphs. That suggests they are looking at how the structure of the signal network changes over time, not just a static snapshot of it.

Lu: The authors propose this dynamic heterogeneous star graph formulation to capture how the topology and link attributes joint capture spatial and temporal dependencies, which is what makes it different from previous work <ref:2509.14000#pg2>.

Meng: So they are building a system that understands the evolving structure of the satellite constellation as it relates to our receiver during jamming events. That’s a deep level of modeling for this kind of problem.

Lalam: It means they are treating the whole signal history, rather than just isolated points, as one big data structure to solve for the positional deviation.

The paper's summary: Tom: The authors summarize JaGNet by saying they reformulate jamming mitigation as a dynamic temporal graph regression problem. They model the satellite-receiver scene at every one Hertz epoch as a dynamic heterogeneous star graph to predict the two dimensional positional deviation caused by jamming.

Jane: It seems like they’re trying to capture both how things are spatially related at any given moment, like signal strength or angles, and how those relationships change over time. It’s not just one-off measurements; it's about the history of the constellation interacting with the receiver <ref:2509.14000#pg2>.

Lu: Exactly. They use a single layer called a Heterogeneous Graph ConvLSTM to fuse that spatial context—things like signal-to-noise ratio, azimuth, and elevation—with those short-term temporal dynamics to predict how far off the position is going to be. It’s modeling the sequence of events in the graph structure itself.

Meng: That makes sense because current methods often look at data sets without really exploiting that spatio-temporal coherence of the constellation. This formulation seems designed specifically to address that weakness you mentioned earlier, Tom.

Lalam: So instead of just correcting a position based on one measurement, JaGuard looks at the whole sequence of interactions to get a better guess about the error trajectory. That’s a big step in how we use these kinds of models for tracking things.

The paper's improvements: Tom: Now, where they really show off their work is how this approach performs compared to existing methods. They test it on an open real-world jamming PEC dataset from two commercial receivers subjected to three different types of interference across six power levels, from negative forty-five decibels down to negative seventy decibels.

Jane: The results are quite compelling, especially under severe jamming conditions. Under the strongest interference at negative forty-five decibels, JaGuard consistently yielded the lowest Mean Absolute Error of twenty-eight point five six centimeters to fifty-nine point two centimeters depending on the jammer type <ref:2509.14000#pg2>.

Lu: That’s a significant reduction when you compare it to advanced time-series and spatial baselines that are typically used for this kind of task. The authors say this model manages to maintain an MAE of twenty-eight point five six centimeters under severe jamming, which is way better than the next best baseline they tested over thirty-five centimeters <ref:2509.14000#pg2>.

Meng: And they even show improvement at lower interference levels; the error drops below two centimeters when you look at weaker interference, which suggests it’s not just good in the worst case. That’s practical for everyday use where jamming might be less intense <ref:2509.14000#pg2>.

Lalam: I was particularly interested in how stable it was when they tested data starvation, where they only had ten percent of the training data. JaGuard kept its error bound between fifteen and twenty centimeters, which is way better than the massive variance increase we saw in their strongest baseline <ref:2509.14000#pg2>.

Conclusion: Tom: So, to wrap up this paper on "JaGNet: Detection, Identification and Position Error Correction of GNSS Jamming with Deep Temporal Graphs," the main point is that by treating jamming mitigation as a dynamic temporal graph regression problem, they create a system that captures spatial and temporal dependencies better than previous methods. It works well even when the training data is scarce.

Jane: Yes, it’s about moving away from static correction methods to something that understands the evolving interaction between the satellites and the receiver over time. It shows how deep temporal graph networks can be used for real-world signal estimation challenges <ref:2509.14000#pg2>.

Lu: The generalization they found, where models trained on one receiver can accurately predict errors on a different receiver under different jamming signatures, is pretty powerful for future deployment scenarios that aren't fully tested in the training data <ref:2509.14000#pg2>.

Meng: From an engineering perspective, the inference time is also low; it takes about one hundred forty-three point seven one milliseconds on a CPU and twenty-two point five six milliseconds on a GPU, which means it doesn't create a major processing bottleneck for real-time applications <ref:2509.14000#pg2>.

Lalam: And the zero shot cross device transferability they demonstrated means we don’t have to train a new model for every single receiver we deploy; one good model can adapt to many situations <ref:2509.14000#pg2>.

Tom: So, JaGNet shows that this deep temporal graph approach is a solid way to tackle GNSS jamming errors using only standard observables. It’s a lot of information for those of you who want to understand how these models are being developed right now.

Department of Communication Systems, Jozef Stefan Institute · Faculty of Electrical Engineering, University of Ljubljana

cs.LG

Submitted: 2025-09-17

Updated: 2026-10-13

Comments: 18 pages, 7 figures

Code: https://github.com/sensorlab/JaGuard

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 88/100

The gist: The gist: JaGuard proposes receiver-centric deep temporal graph networks to correct position errors caused by GNSS jamming by reformulating jamming mitigation as a dynamic temporal graph regression

Key concepts

Dynamic Heterogeneous Graph Regression: This formulation treats the satellite-receiver interaction as a sequence of graphs that change over time. It captures both the spatial relationships (who is visible) and the temporal evolution (how visibility and signal quality change) simultaneously to predict positional errors.
HeteroGCLSTM: A single-layer deep learning model combining Graph Convolutional Networks with Long Short-Term Memory networks. It processes heterogeneous data—features from different node types like the receiver and satellites—to effectively capture both spatial structure and short-term temporal dependencies in the GNSS signal.
Receiver-Centric Star Graph: The model represents the satellite constellation as a star graph where the receiver is the central node connected to visible satellites. This structure is chosen because it naturally models how a single receiver aggregates information from its immediate neighbors, which is crucial for tracking rapid changes during jamming.

Terminology

Global Navigation Satellite Systems — 1× in this paper
Heterogeneous Graph — 1× in this paper · explained in 2 episodes
Mean Absolute Error — 1× in this paper

Summary

The gist: JaGuard proposes receiver-centric deep temporal graph networks to correct position errors caused by GNSS jamming by reformulating jamming mitigation as a dynamic temporal graph regression problem.

Problem Statement

Global Navigation Satellite Systems (GNSS) face growing disruption from low-cost intentional jammers, which cause positional drifts of tens of centimeters to metres under interference. Current Position error correction (PEC) works only focus on datasets containing errors due to multi-path propagation and do not exploit the spatio-temporal coherence of satellite constellations. The fundamental vulnerability stems from the extremely weak received GNSS signal power, further degraded by satellite aging. Jamming remains the most operationally disruptive mechanism. Existing position error correction methods remain static, relying on single-epoch observations.

Proposed Methodology

The paper reformulates jamming-induced PEC as a dynamic heterogeneous graph regression problem. At each 1 Hz epoch, the satellite–receiver scene is modeled as a heterogeneous star graph. A single-layer Heterogeneous Graph ConvLSTM (HeteroGCLSTM) fuses one-hop spatial context (Signal-to-Noise ratio (SNR), azimuth, elevation) with short-term temporal dynamics to predict the 2D positional deviation. The model operates exclusively on standard NMEA observables (SNR, azimuth, elevation, position), requiring no raw pseudorange or carrier-phase data. The architecture employs a single Heterogeneous Graph Convolutional Long Short-Term Memory (HeteroGCLSTM) layer [38], a choice justified by the star-graph topology.

Evaluation and Results

JaGuard was evaluated using an open real-world jamming PEC dataset from two commercial receivers subjected to highly controlled, synthesized RF interference across three jammer types and six power levels (−45 to −70 dBm) Compared to advanced time-series and spatial baselines, our model consistently yield the lowest Mean Absolute Error (MAE) Under severe jamming (−45 dBm), it maintains an MAE of 2.85–5.92 cm, improving to sub-2 cm accuracy at lower interference levels. On mixed-power datasets, JaGuard substantially outperforms all baselines with MAEs of 2.26 cm (Ai-Thinker GP01) and 2.61 cm (U-blox MAX-M10) Under extreme data starvation (10% training data), JaGuard remains highly stable, bounding the error at 15–20 cm and preventing the massive variance increase observed in the strongest baseline.

Comparative Analysis

JaGuard consistently achieves the lowest errors across all jamming types and power levels. At −45 dBm, it yields 3.56–4.03 cm depending on jammer type, roughly halving the next-best baseline. The spatial GCNN fails with errors exceeding > 35 cm 35 cm>. JaGuard’s node-level spatiotemporal modeling captures individual signal degradation trajectories that global set-based or flattened models cannot resolve as precisely.

Operational Feasibility

JaGuard achieves a mean inference time of 143.71 ms on CPU (AMD EPYC 7543P) and 22.56 ms on GPU (NVIDIA L40S), consuming less than 15% of the available 1 s time budget even without GPU acceleration. JaGuard eliminates this bottleneck by encoding both spatial topology and temporal dynamics within a single heterogeneous messagepassing step, reducing peak GPU memory to 17.41 MB.

Conclusion

JaGuard demonstrates exceptional data efficiency, tightly bounding errors even with only 10% training data – a regime in which baselines degrade severely. Furthermore, it exhibits robust zero-shot cross-device transferability, with models trained on the U-blox 10 generalizing accurately to the GP01 (1.59 cm MAE) and vice versa (3.18 cm MAE) under unseen jamming signatures. Future work will extend JaGuard to kinematic mobile deployments, investigate ondevice inference on embedded mobile platforms, and investigate theoretical conditions under which resilient estimation remains feasible.

How it works

The satellite–receiver scene is modeled as a heterogeneous star graph at each 1 Hz epoch.
A single-layer Heterogeneous Graph ConvLSTM (HeteroGCLSTM) fuses one-hop spatial context (Signal-to-Noise ratio (SNR), azimuth, elevation) with short-term temporal dynamics to predict the 2D positional deviation.
The model unrolls through k snapshots, mapping inputs into a 256-dimensional hidden space <ref:4pg4, As shown in Fig.

Improvements for AI systems

Based on JaGuard, a receiver-centric deep temporal graph network, an improved system can estimate 2D positional deviation by modeling the satellite–receiver interaction as a sequence of heterogeneous star graphs to capture both spatial and temporal dependencies.
The JaGuard architecture, specifically the HeteroGCLSTM layer, allows the system to process sequences of dynamic graph snapshots where node sets and edge structures change between epochs, enabling it to model the rapid disruption of spatiotemporal coherence caused by broadband jamming.
The improved system can operate exclusively on standard NMEA observables, as the formulation relies exclusively on standard NMEA observables (SNR, azimuth, elevation, position), requiring no raw pseudorange or carrier-phase data, ensuring deployment on any off-the-shelf receiver.
The model provides robust performance under data constraints by demonstrating that under extreme data starvation (10% training data), JaGuard remains highly stable, bounding the error at 15–20 cm and preventing the massive variance increase observed in the strongest baseline.
The system achieves strong generalization across hardware by exhibiting robust zero-shot cross-device transferability, where models trained on one receiver can generalize to an unseen receiver under different jamming signatures, such as the GP01-trained model encounters unseen FM jamming on U-blox 10.

Sources

TSMixer: An All-MLP Architecture for Time Series Forecasting