Session G2b: Resilient and Secure PNT Against Interference and Emerging Threats 1

Cross-Domain Feature Fusion for GNSS Spoofing Detection with Lightweight Spatio-Temporal Attention Networks
Jiale Ye, Xin Chen, School of Automation and Intelligent Sensing, Shanghai Jiao Tong University
Location: Palm

Global Navigation Satellite System (GNSS) serves as the core infrastructure for positioning, navigation, and timing (PNT) in safety-critical applications, such as autonomous driving and unmanned aerial vehicles (UAVs). However, civilian GNSS signals are publicly modulated and extremely weak when propagating to the ground surface, rendering them highly vulnerable to spoofing attacks and multipath interference. Such interference can degrade positioning accuracy or even mislead receivers into outputting falsified information, thereby posing serious threats to navigation safety.
Existing GNSS interference detection techniques can be primarily categorized into two types: traditional signal processing approaches and deep learning-based intelligent methods. Traditional methods predominantly construct test statistics based on signal features from a single domain, which inherently suffer from limitations including difficulties in threshold determination, insufficient detection sensitivity for covert spoofing, and poor robustness in complex environments. Meanwhile, mainstream deep learning-based detection methods generally face three core bottlenecks: the scarcity of high-quality labeled datasets, inadequate mining of multi-domain signal feature information leading to limited detection discriminability, and high model complexity that impedes real-time deployment on edge devices.
To address these limitations, this paper proposes three key strategies. First, we release a measured GNSS spoofing and multipath dataset encompassing diverse scenarios, interference types, and motion states, which fills the gap in high-fidelity labeled data for related research. Second, we construct a cross-domain feature vector system that fuses the correlation, observation, and positioning domains of GNSS signals, enabling multi-scale and multi-granularity mining of intrinsic differences between spoofing and authentic signals. Third, we design a Lightweight Spatio-Temporal Attention Network (LSTAN) tailored to the structural characteristics of GNSS cross-domain high-dimensional tensors and the temporal evolution law of spoofing attacks, which achieves a balance between high detection accuracy, low false alarm rate, and short inference latency.
In the correlation domain, we extract Signal Quality Monitoring (SQM) indicators, including Ratio, Delta, and ELP, from correlator outputs at the MHz-level sampling rate to capture fine-grained anomalies in the tracking loop. These metrics quantify the distortion of correlation peaks in terms of code symmetry, energy distribution, and carrier phase stability, thereby forming single-satellite correlation feature vectors.
In the observation domain, we derive Hz-level satellite measurements, including pseudorange, carrier phase, Doppler shift, and carrier-to-noise ratio (C/N?). Two types of features are extracted: signal power temporal distribution, which capture abrupt power fluctuations, and single-satellite physical consistency features. The latter are constructed based on the inherent relationships among pseudorange rate, carrier phase rate, and Doppler-derived velocity, reflecting the violations of physical laws under spoofing attacks.
In the positioning domain, we extract receiver-level global features, covering pseudorange residual distribution, clock bias variation rationality, and position trajectory continuity. Authentic signals typically yield zero-mean Gaussian residuals and smooth clock variations, while spoofing attacks introduce systematic biases and abnormal jumps in these parameters. Trajectory continuity features capture gradual or abrupt deviations from normal motion patterns.
To fuse these heterogeneous domain features, we design a fusion strategy: time-axis alignment using the low-frequency positioning domain as a benchmark; feature-granularity mapping that broadcasts positioning-domain global features to each satellite dimension; and high-dimensional tensor assembly that integrates all features into a unified tensor. The tensor is normalized using the z-score method to eliminate dimensional differences, forming a physically interpretable and highly discriminative spoofing fingerprint.
The LSTAN network takes this fused tensor as input and cascades depthwise separable convolution, bidirectional gated recurrent unit (BiGRU), and an attention module, combined with residual connections and layer normalization. Depthwise separable convolution encodes multi-scale spatial correlations while reducing the number of model parameters by more than 60% compared to standard convolution. The BiGRU captures long-range temporal dependencies with a moderate hidden dimension, outputting a spatio-temporal feature sequence. The attention module first applies global average pooling and max pooling, then generates channel attention weights through a lightweight two-layer fully connected network, performing element-wise weighting to enhance discriminative features and suppress noise. It subsequently computes temporal attention weights via linear transformation and Softmax function, aggregating the feature sequence into a global context vector. Finally, a lightweight classification head with layer normalization and Sigmoid activation outputs the spoofing probability.
Experimental evaluations are conducted using both the public TEXBAT dataset and our self-collected dataset, which is recorded with a spoofing generator. The generator adopts a capture-modify-regenerate architecture. It receives live GNSS signals, decodes navigation messages, adjusts the transmission time, and broadcasts spoofing signals via HackRF. Our dataset covers both static and dynamic scenarios and includes multipath samples collected from urban canyons.
Comparative experiments on TEXBAT subsets demonstrate that ResNet18 achieves an accuracy of 90.78% with 1.71 million parameters, while LSTM achieves 97.48% accuracy with 1.16 million parameters. In contrast, the proposed LSTAN achieves 96.65% accuracy with only 0.55 million parameters—less than half the parameter count of LSTM—exhibiting a superior lightweight design while maintaining competitive detection performance. The limited sample diversity of the TEXBAT dataset poses generalization risks. To address this, we incorporate data from INF-DEV-01 to enrich scenario coverage and enhance the model's robustness for real-world deployment.
Based on these experimental results, the proposed method is expected to achieve a detection probability of no less than 99%, a false alarm rate of no higher than 1%, and an end-to-end response latency of less than 3 seconds. Future work will extend the framework to our larger self-constructed dataset, conduct comparative analyses with traditional machine learning methods, and validate the feasibility of rapid positioning recovery through interference isolation strategies.