Scholarlink

<p>Global Navigation Satellite System (GNSS) spoofing attacks pose significant threats to economic and public security, necessitating rapid and accurate detection methods. While machine learning-based approaches have surpassed traditional rule-based statistical methods by leveraging multi-domain information fusion, they primarily&#xa0;learn data distributions rather than intrinsic signal invariants, resulting in degraded performance under out-of-distribution scenarios.To address this limitation this paper proposes a Code-Carrier Consistency Physics-Constrained Neural Networks (CCC-PCNNs) for spoofing detection. The proposed method exploits the inherent code-carrier consistency of GNSS signals by modeling their first-order differential relationship as a physical constraint. This constraint is incorporated into the loss function of a one-dimensional convolutional neural network (1D-CNN), enabling the model to adhere to underlying physical laws while optimizing detection performance. Unlike conventional approaches that rely solely on feature engineering, the proposed framework embeds domain knowledge directly into the learning process, thereby enhancing generalization and robustness. Experimental results on three&#xa0;widely used open-source datasets, Texas Spoofing Test Battery (TEXBAT), Oak Ridge Spoofing and Interference Test Battery (OAKBAT), and Finnish Geospatial Research Institute Spoofing Repository (FGI-SpoofRepo), demonstrate the proposed algorithm achieves an average detection accuracy exceeding 98.00% under within-scenario supervised evaluation. With the incorporation of only a single additional feature beyond the baseline code-carrier phases, the proposed algorithm achieves an average improvement of 10.42% over the baseline 1D-CNN across all three datasets. More significantly, systematic evaluation using the Accuracy Retention Rate (ARR) metric shows strong cross-scenario and cross-dataset generazation.Across 10 unseen scenarios from the TEXBAT and OAKBAT datasets, the CCC-PCNNs model maintains ARR values consistently above 92.70%, achieving improvements of 23.49%, 18.40%, 9.39%, 28.82%, 24.19%, 26.99% and 46.16% over the 1D-CNN, Long Short-Term Memory (LSTM), Transformer, Support Vector Machine (SVM), Residual Network (Resnet), Graph Attention Networks (GAT) and CCC-Detector models, substantially outperforming traditional machine learning approaches. This research innovatively integrates physical mechanisms with data-driven methods to overcome generalization challenges arising from the variability and unpredictability of spoofing attacks in adversarial environments, strengthening the security of critical infrastructure and core systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Generalized GNSS spoofing detection in unseen scenarios using physics-constrained neural networks with code-carrier consistency

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月26日

查看全文
  • Yili Song,
  • Wei Xiao,
  • Xiaozhou Ye,
  • Xin Yang,
  • Wenxiang Liu,
  • Guangfu Sun

摘要

Global Navigation Satellite System (GNSS) spoofing attacks pose significant threats to economic and public security, necessitating rapid and accurate detection methods. While machine learning-based approaches have surpassed traditional rule-based statistical methods by leveraging multi-domain information fusion, they primarily learn data distributions rather than intrinsic signal invariants, resulting in degraded performance under out-of-distribution scenarios.To address this limitation this paper proposes a Code-Carrier Consistency Physics-Constrained Neural Networks (CCC-PCNNs) for spoofing detection. The proposed method exploits the inherent code-carrier consistency of GNSS signals by modeling their first-order differential relationship as a physical constraint. This constraint is incorporated into the loss function of a one-dimensional convolutional neural network (1D-CNN), enabling the model to adhere to underlying physical laws while optimizing detection performance. Unlike conventional approaches that rely solely on feature engineering, the proposed framework embeds domain knowledge directly into the learning process, thereby enhancing generalization and robustness. Experimental results on three widely used open-source datasets, Texas Spoofing Test Battery (TEXBAT), Oak Ridge Spoofing and Interference Test Battery (OAKBAT), and Finnish Geospatial Research Institute Spoofing Repository (FGI-SpoofRepo), demonstrate the proposed algorithm achieves an average detection accuracy exceeding 98.00% under within-scenario supervised evaluation. With the incorporation of only a single additional feature beyond the baseline code-carrier phases, the proposed algorithm achieves an average improvement of 10.42% over the baseline 1D-CNN across all three datasets. More significantly, systematic evaluation using the Accuracy Retention Rate (ARR) metric shows strong cross-scenario and cross-dataset generazation.Across 10 unseen scenarios from the TEXBAT and OAKBAT datasets, the CCC-PCNNs model maintains ARR values consistently above 92.70%, achieving improvements of 23.49%, 18.40%, 9.39%, 28.82%, 24.19%, 26.99% and 46.16% over the 1D-CNN, Long Short-Term Memory (LSTM), Transformer, Support Vector Machine (SVM), Residual Network (Resnet), Graph Attention Networks (GAT) and CCC-Detector models, substantially outperforming traditional machine learning approaches. This research innovatively integrates physical mechanisms with data-driven methods to overcome generalization challenges arising from the variability and unpredictability of spoofing attacks in adversarial environments, strengthening the security of critical infrastructure and core systems.