⚡ electrical engineering

HAS-Net: Hierarchical Cross-PRN Attention for Single-Antenna GNSS Spoofing Detection

This paper introduces HAS-Net, a novel hierarchical attention network that fuses multi-correlator features across all visible satellites to achieve robust single-antenna detection of time- and phase-aligned matched-power GNSS spoofing attacks, significantly outperforming both classical and existing machine learning baselines.

Original authors: Seungsoo Yoo, Sun Yong Kim

Published 2026-10-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Seungsoo Yoo, Sun Yong Kim

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Global Positioning System is the invisible grid that guides everything from the smartphone in your pocket to the navigation systems of massive cargo ships. It works by listening to faint radio signals beamed from a constellation of satellites orbiting high above the Earth. Because these signals are unencrypted and their structure is public knowledge, they are vulnerable to a specific kind of deception known as spoofing. In a spoofing attack, a malicious actor broadcasts a fake signal that mimics the real one, tricking a receiver into believing it is somewhere else or that time has shifted. While it is relatively easy to spot a crude fake signal that is louder or out of sync with the real satellites, the most dangerous attacks are those that are perfectly synchronized. These "matched-power" attacks arrive with the exact same strength and timing as the genuine signals, making them nearly invisible to standard detection methods that rely on a single antenna. For years, this has been an open problem in security: how do you tell the difference between a real satellite and a perfect copy when you only have one ear to listen with?

Researchers at Konkuk University have developed a new approach to solve this puzzle, moving beyond simple checks to a system that learns to spot the subtle inconsistencies that only appear when multiple satellites are considered together. They created a tool called HAS-Net, which stands for Hierarchical Cross-PRN Attention Network. Instead of looking at each satellite signal in isolation, the system treats the entire group of visible satellites as a single, interconnected team. It uses a type of artificial intelligence known as deep learning to analyze the tiny distortions in the radio waves that occur when a fake signal tries to take over. The key insight is that while a sophisticated attacker might perfectly mimic one or two satellites, they often struggle to maintain that perfect mimicry across all the satellites a receiver can see at once. By comparing the signals from all the satellites against each other, the system can detect the slight mismatch that reveals the deception.

To build and test this system, the researchers used a sophisticated software receiver that acts like a high-speed digital ear, capable of breaking down radio signals into incredibly fine details. They fed the system data from a well-known collection of test scenarios called TEXBAT, which includes recordings of various types of spoofing attacks, ranging from simple disruptions to the most advanced, perfectly synchronized fakes. The system was trained on five different scenarios and then tested on two it had never seen before, including the most difficult type of attack where the fake signal is perfectly aligned in both time and phase with the real one. The results were striking. In these hardest tests, traditional methods that rely on simple mathematical thresholds failed completely, often missing the attack entirely. Even more advanced machine learning models that looked at each satellite individually struggled, missing the attack in a significant number of cases during the critical first twenty seconds after the fake signal turned on.

The new system, however, achieved a perfect record in detecting these attacks the moment they began. When the researchers tested it against the most challenging, perfectly synchronized fake signal, the system identified the deception in every single instance during the first twenty seconds of the attack. This is a crucial distinction because in a real-world scenario, every second of delay allows the fake signal to pull the navigation system further off course, potentially leading a vehicle or ship into danger. The researchers found that the secret to this success was not just the complexity of the computer model, but specifically the way it compared the satellites. When they removed the part of the system that allowed the satellites to "talk" to each other, the detection rate dropped dramatically. This proved that the ability to fuse information across the entire group of satellites is the missing piece that makes single-antenna detection possible.

The study also highlighted a limitation: the system was trained on data from a stationary receiver. When the researchers tested it on data involving a moving receiver, the performance dropped, suggesting that the motion of the vehicle introduces new patterns that the system has not yet learned to ignore. This indicates that while the method is a major step forward, it is not yet a universal solution for all moving vehicles. Nevertheless, the findings demonstrate that it is possible to detect even the most sophisticated, perfectly synchronized spoofing attacks using only a single antenna. By shifting the focus from checking individual signals to understanding the relationships between them, the researchers have opened a new path for securing the global navigation systems that modern society relies upon. The work suggests that the future of GPS security may lie not in adding more hardware, but in teaching receivers to think more holistically about the signals they receive.

Technical Summary: HAS-Net for Single-Antenna GNSS Spoofing Detection

Problem Statement
Detecting time- and phase-aligned, matched-power Global Navigation Satellite System (GNSS) spoofing attacks using a single antenna remains a significant open challenge in civilian receiver security. While classical Signal-Quality Monitoring (SQM) techniques (e.g., the $\Delta$ -test) are computationally efficient, they fail to detect sophisticated attacks (specifically TEXBAT scenarios ds7 and ds8) once the spoofing signal has captured the receiver's tracking loops. The distortion of the correlation function in these regimes is too subtle for scalar thresholds to distinguish from noise without unacceptable false-alarm rates. Furthermore, existing machine learning approaches often treat each satellite (Pseudorandom Noise code, or PRN) independently, failing to leverage the critical inconsistency between spoofed and unspoofed satellites within the same scenario.

Methodology: HAS-Net Architecture
The authors propose HAS-Net (Hierarchical Cross-PRN Attention Network), a deep learning framework designed to fuse correlator distortions across all visible satellites. The system utilizes the open-source FGI-GSRx software receiver to extract 11 features per 1 ms epoch from a 17-finger multi-correlator tracker. These features include:

5 SQM metrics: Delta, Ratio, ELP (Early-Late Power), Slope, and MLres (correlation peak residual).
3 Consistency metrics: Code-loop discriminator consistency (CCD), Phase-loop discriminator consistency (DPR), and phase jump ( $\phi_{jump}$ ).
3 Auxiliary metrics: Carrier-to-noise density ( $C/N_0$ ), Automatic Gain Control (AGC) response, and power variance ( $P_{var}$ ).

The HAS-Net architecture processes a 5-second window (500 time steps) of these features for up to 14 PRNs through four stages:

Per-PRN Encoder: A shared bidirectional Long Short-Term Memory (LSTM) network independently encodes the temporal signature of each satellite's feature vector.
Cross-PRN Self-Attention: A multi-head self-attention block (4 heads) pools information across all visible PRNs. This component is permutation-equivariant, allowing the model to learn relationships between satellites without relying on specific satellite identities.
Masked Mean-Pool Aggregator: This layer collapses the attention outputs into a single scenario-level embedding, excluding zero-padded PRN slots via a binary mask.
Classifier Head: A two-layer Multilayer Perceptron (MLP) with ReLU activation and dropout produces the final binary classification (spoof vs. clean).

The model is trained on five TEXBAT scenarios (cleanStatic80, ds1, ds2, ds5, ds7) and evaluated on held-out scenarios ds4 (matched-power, position-push) and ds8 (matched-power, time- and phase-aligned).

Key Contributions

Cross-PRN Attention Mechanism: The paper introduces cross-PRN self-attention as an explicit architectural primitive for single-antenna spoofing detection. An ablation study demonstrates that this block alone contributes a +0.65 absolute recall gain on the ds8 aligned phase compared to a per-PRN BiLSTM baseline, identifying inter-satellite information fusion as the critical missing ingredient for detecting phase-aligned attacks.
Phase-Wise Evaluation Protocol: The authors propose a detection-latency-aware evaluation protocol that breaks down performance into four phases: pre-onset, onset (first 20s), aligned, and drift. They argue that standard Area Under the Curve (AUC) metrics can mask failures in the critical "onset" phase, where early detection is operationally vital.
Reproducible Pipeline: The authors release a fully reproducible pipeline combining the FGI-GSRx receiver (with a non-invasive patch directory), the 11-feature extraction logic, and the HAS-Net training script implemented in MATLAB.

Results
Evaluated on the TEXBAT ds4 and ds8 held-out test sets:

Overall Performance: HAS-Net achieves a scenario-window-level test AUC of 0.925. While Random Forest and k-Nearest Neighbors (kNN) showed higher AUCs (0.999 and 0.974, respectively), they exhibited severe overfitting (train-validation gaps of 25 and 6 points, respectively) and poor generalization to the ds8 attack.
Onset Detection: HAS-Net is the only method to achieve a seed-stable 1.000 onset-phase recall on the ds8 attack (matched-power, phase-aligned) across five independent training seeds. In contrast, the strongest classical baseline ( $\Delta$ -test) achieved 0.0 recall, and the strongest learned baseline (Linear SVM) achieved 0.875 recall. On the ds4 attack, the onset recall averaged 0.92 ± 0.18 across seeds (achieving 1.000 in four of five seeds and 0.600 in one), whereas the ds8 result remained stable at 1.000.
Ablation: Removing the cross-PRN attention block reduced the ds8 aligned recall from 1.000 to 0.350, confirming that the improvement stems from fusing information across satellites rather than increased model capacity alone.
Limitations: The model shows poor generalization to moving receivers (the cleanDynamic validation scenario), suggesting that receiver-motion artifacts are out-of-distribution for a model trained on static data. Additionally, the model assumes the attacker is a known class (trained on ds7); generalization to entirely unseen attack classes remains an open problem.

Significance
The paper claims that HAS-Net closes the gap in single-antenna spoofing detection by demonstrating that inter-satellite inconsistency is the detectable signature of phase-aligned, matched-power spoofing. By leveraging a hierarchical attention mechanism, the system can detect attacks within the first 20 seconds of activation (the onset phase) with zero miss-rate on the most challenging TEXBAT scenarios (specifically ds8), a capability that neither classical SQM nor standard per-PRN machine learning classifiers possess. The work establishes that fusing signal-quality information across the visible constellation is the necessary step to secure civilian receivers against advanced spoofing without requiring additional hardware.

Technical Summary: HAS-Net for Single-Antenna GNSS Spoofing Detection

More like this