⚡ electrical engineering

Physics-Informed Dual-Modality Monitoring for GNSS Spoofing Detection

This paper introduces Physics-Informed Dual-Modality Cross-Attention Networks (PI-DMCANs), a novel deep learning framework that fuses tracking features and cross-ambiguity function images with physics-based residual gating to achieve superior GNSS spoofing detection accuracy across diverse scenarios compared to existing baselines.

Original authors: Liupeng Chen, Gang Ou, Zhen Huang, Zheng Wei, Weihua Mou

Published 2026-10-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Liupeng Chen, Gang Ou, Zhen Huang, Zheng Wei, Weihua Mou

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 navigation satellite systems that guide our cars, ships, and phones rely on signals that are incredibly weak by the time they reach the ground. Because these signals are open and unencrypted, they are vulnerable to a specific kind of deception known as spoofing. In a spoofing attack, a bad actor broadcasts a fake signal that mimics the real one, tricking a receiver into believing it is somewhere it is not or that time is passing at the wrong speed. Unlike simple jamming, which just blocks the signal and leaves the user with no position, spoofing is more dangerous because it provides a false position that looks perfectly normal. The receiver continues to work, unaware that it has been hijacked, which can lead to catastrophic errors in critical infrastructure, emergency response, or autonomous driving.

To protect against this, engineers have long tried to spot the subtle differences between a real satellite signal and a fake one. They look at how the signal behaves as it is tracked over time and examine the shape of the signal's peak when it is first captured. However, these methods often struggle when the type of attack changes or when the receiver is moving. A method that works well against one kind of fake signal might fail completely against another, especially if the attacker is clever enough to make their fake signal look very similar to the real thing. The challenge is to build a system that can recognize these deceptions reliably, even when it has never seen that specific type of attack before.

A team of researchers at the National University of Defense Technology has developed a new approach to solve this problem, described in a recent study. They created a monitoring system that acts like a dual-sense observer, looking at the satellite signal through two different lenses simultaneously. The first lens watches a short sequence of how the signal is being tracked, noting small fluctuations in power and timing. The second lens looks at a detailed image of the signal's shape, capturing how the signal spreads out over time and delay. By combining these two views, the system can see patterns that a single view would miss.

The researchers built a computer model that fuses these two streams of information using a technique called cross-attention. This allows the system to decide, for every single satellite it sees, which view is more trustworthy at any given moment. If the signal looks suspicious in the tracking data but normal in the image, the system weighs the tracking data more heavily. If the image shows a strange split or distortion that the tracking data misses, the system shifts its focus to the image. This dynamic weighting is guided by a set of physical rules that the researchers programmed into the system. These rules check for specific inconsistencies, such as whether the signal's timing matches its speed or whether the signal's strength is behaving in a way that only a fake signal would.

To test their creation, the researchers used three different collections of real-world and simulated data, including standard datasets used by the scientific community and a new, multi-constellation dataset they helped create. They compared their new system against five existing methods, ranging from simple statistical tests to more complex neural networks. In tests where the system was evaluated on the same type of attack it had been trained on, it achieved an accuracy of 96.5 percent. More importantly, when the researchers tested it on completely new types of attacks that it had never seen before, the system maintained a high level of performance, achieving an accuracy of 93.2 percent. This was significantly better than the next best method, which dropped to 88.1 percent under the same conditions.

The study explicitly rules out the idea that a single type of measurement is enough to catch all spoofing attacks. The researchers found that relying only on the signal's image or only on its tracking history leads to failures when the attack method changes. They also demonstrated that simply combining the two data streams with equal weight is not as effective as letting the system decide which stream to trust based on physical evidence. The results show that by letting the system adapt its focus based on the specific physical behavior of the signal, it becomes much harder for a spoofing attack to go undetected.

The practical outcome of this work is a tool that can be installed in a standard GPS receiver to flag specific satellites that are likely being spoofed. Once a satellite is flagged, the receiver can ignore its data or reduce its influence on the final position calculation, allowing the system to continue working correctly even while under attack. This is particularly valuable for users who do not have access to expensive equipment like multiple antennas or inertial navigation systems. The researchers confirmed that their method works across different types of satellite systems, including GPS, Galileo, and BeiDou, making it a versatile solution for the modern, multi-satellite world. While the system is not perfect and still faces challenges with certain complex, synchronized attacks, it represents a significant step forward in making our navigation systems more resilient to deception.

Technical Summary: Physics-Informed Dual-Modality Monitoring for GNSS Spoofing Detection

Problem Statement
Global Navigation Satellite Systems (GNSS) are vulnerable to spoofing, where fabricated signals manipulate receiver tracking loops onto a prescribed trajectory without necessarily denying the fix. While multi-constellation redundancy improves geometry, it does not inherently mitigate radio-frequency threats. Existing detection methods face three primary challenges:

Heterogeneity and Asynchrony: Tracking features (time-series) and correlator images (spatial-temporal heatmaps) operate on different time scales and have distinct missingness patterns, making simple concatenation ineffective.
Scenario Dependency: Single-metric monitors (e.g., Signal Quality Monitoring or SQM) or shallow learning models often fail when the attack type (e.g., matched-power vs. overpowered) or the constellation changes, as they memorize scenario-specific fingerprints rather than transferable physical structures.
Generalization Limits: Current physics-informed approaches often rely on a single physical constraint (e.g., code-carrier consistency) which may be insensitive to specific attack phases (e.g., the middle of a matched-power interval) or may inflate false alarms on clean channels with weak antennas or multipath.

The paper addresses the need for a channel-wise monitor that can identify spoofed satellites on a single-antenna receiver without inertial aids, operating effectively across unseen scenarios and constellations.

Methodology: PI-DMCANs
The authors propose Physics-Informed Dual-Modality Cross-Attention Networks (PI-DMCANs). The architecture fuses two distinct modalities using causal cross-attention, with the fusion weights dynamically conditioned by a bank of physical residuals.

Dual-Modality Observations:
- Tracking Features: A short sequence (1s window) of six tracking-loop statistics: Ratio, Delta, ELP (SQM metrics), $C/N_0$ , $C/N_0$ standard deviation, and Code-Minus-Carrier (CMC) consistency.
- CCAF Image: A local magnitude and quadrature (Mag/Q) heatmap of the Cross-Ambiguity Function (CCAF) over a $\pm 10$ chip code-delay neighborhood. This image is peak-normalized to remove absolute power dependence while preserving secondary peak geometry.
Encoders:
- Tracking Encoder: A bidirectional LSTM processes the tracking sequence to capture non-stationary transients around takeover events.
- CCAF Encoder: A 2D CNN extracts frame-level geometry from the Mag/Q heatmap, followed by a unidirectional GRU to aggregate temporal evolution causally.
Causal Cross-Attention:
- Tracking tokens serve as queries, while CCAF tokens serve as keys and values. This mechanism aligns the CCAF context to the tracking time axis, allowing the network to retrieve relevant image features based on the tracking state.
Physics-Informed Residual Gating:
- A Residual Bank computes five physical test statistics relative to clean references: Code-Carrier Consistency ( $r_{ccc}$ ), SQM level/jitter ( $r_{sqm}$ ), Peak Walk ( $r_{warp}$ ), Secondary Peak ( $r_{twin}$ ), and Clock Drift ( $r_{clk}$ ).
- These residuals are processed by a gating network (MLP) with a learnable soft physical prior. The gate outputs weights ( $g_{trk}, g_{img}$ ) that determine the contribution of each modality. For instance, if the SQM residual is high, the gate may prioritize the tracking stream; if peak splitting is detected in the image, the CCAF stream is weighted higher.
- This gating ensures that the less informative stream does not mislead the decision, particularly under hold-out conditions.
Training Objective:
- The model is trained with a composite loss: a focal detection loss (weighted for spoofing takeover transients), a gate regularizer (entropy penalty to prevent gate collapse), a channel contrastive loss (InfoNCE) to distinguish physical states of different satellites, and a one-sided score-residual consistency term. This consistency term ensures the detection score aligns with large observed residuals without forcing residuals to zero (which would penalize clean multipath).

Key Contributions

Dual-Modality Cross-Attention: A novel fusion mechanism that treats tracking sequences and CCAF images as heterogeneous streams, using causal attention to align them rather than simple concatenation.
Residual-Conditioned Gating: A mechanism where a bank of physical residuals dynamically weights the modalities. This allows the system to adaptively rely on the most informative modality for a specific attack type (e.g., relying on tracking features when peak shapes are ambiguous).
Cross-Constellation Generalization: The method converts residuals to range units using constellation-specific reference frequencies, allowing a shared detector to operate across GPS, Galileo, and BeiDou while maintaining constellation-specific thresholds for clean baselines.

Results
The method was evaluated on three datasets: TEXBAT, OAKBAT, and the multi-constellation MuCoGS, using a Neyman-Pearson (NP) threshold based on the 95th percentile of clean training scores.

Within-Scenario: PI-DMCANs achieved an accuracy of 0.965, outperforming the strongest baseline (CCC-PCNNs at 0.941) and image-only models (CAF-CNN at 0.940).
Within-Dataset Cross-Scenario: When one scenario was held out for testing, PI-DMCANs achieved 0.928 accuracy, approximately 3.5 percentage points higher than the strongest baseline (CAF-CNN at 0.893).
Cross-Dataset Evaluation: In a rigorous test where the held-out scenario was unseen and training data came from different datasets, PI-DMCANs achieved 0.932 accuracy, roughly 5.1 percentage points above the strongest baseline (CAF-CNN at 0.881).
Ablation Studies: Removing the residual gating (Equal-fuse) reduced cross-dataset accuracy to 0.909. Removing either modality caused significant drops (Trk-only: 0.860; CCAF-only: 0.821), confirming that both streams and the gating mechanism are essential for generalization.

Significance and Claims
The paper claims that PI-DMCANs provides a robust, channel-wise spoofing detection mechanism that does not require multi-antenna arrays or inertial units. By fusing heterogeneous modalities through physics-informed gating, the system maintains high detection rates even when the attack scenario or constellation is unseen during training.

The authors emphasize that the resulting "channel flag" serves as a direct basis for interference mitigation, allowing a receiver to de-weight or exclude anomalous satellites before the navigation solution is computed. The work highlights that while single-metric or single-modality approaches struggle with matched-power spoofing and cross-constellation generalization, the proposed dual-modality approach with residual conditioning successfully transfers physical structure across diverse attack families and datasets. The authors note limitations regarding common-mode clock drift attacks (which require external time references) and the sensitivity of the NP threshold to score scale shifts in completely new scenarios.

Technical Summary: Physics-Informed Dual-Modality Monitoring for GNSS Spoofing Detection

More like this