Session A2: GNSS Security: Interference, Jamming, and Spoofing 1

Spoofing Data Augmentation Based on GAN
Lei Xu, Chao Sun, Lu Bai, Shuai Zhang, Yanhong Kou, Ying Xu, Beihang University
Location: Beacon A

Global Navigation Satellite Systems (GNSS) have been extensively applied in various fields; however, their signals are susceptible to interference, particularly spoofing interference, which is stealthy and highly effective. When a GNSS receiver picks up spoofing interference signals, it mistakenly interprets these as legitimate satellite transmissions, leading to erroneous position calculations that can misdirect vehicles, ships, or aircraft, potentially causing severe safety incidents. Due to the rarity of spoofing interference events and the sensitivity of associated data, coupled with the inability of simulation environments to fully replicate the complexity and variability of the real world, research in this area faces challenges due to insufficient data. This limitation hampers the training effectiveness and generalization capability of models.
Data augmentation techniques, which generate more diversified samples by transforming existing data, can expand the scale and diversity of datasets, thereby improving model performance when faced with different types of spoofing interference. Since the introduction of Generative Adversarial Networks (GANs) by Ian Goodfellow et al. in 2014, and the subsequent development of the Wasserstein GAN (WGAN) by Gulrajani et al. in 2017, these methods have been further refined to improve the stability and quality of generated samples. In 2021, Demir S et al. proposed a combination of autoencoders (AE), variational autoencoders (VAE), and WGAN with gradient penalty for time series data augmentation. However, current methods mainly target areas such as image processing, where geometric transformations, color changes, noise addition, and synthetic operations are used, and they do not suitably apply to GNSS signal data.
To address the limitations of traditional data augmentation methods under small sample conditions, this paper extends the WGAN to the area of GNSS spoofing detection. This approach effectively handles the problem of limited sample size to achieve data augmentation for satellite navigation data. Compared to traditional GAN methods, the WGAN method improves the quality and detection rate of generated data through Wasserstein distance and gradient penalty optimized GAN architecture. The main contributions of this paper are summarized as follows:
1. Data Augmentation Using GAN. By employing an adversarial training process between generator and discriminator networks to learn data distributions, the generator creates realistic samples while the discriminator attempts to distinguish authentic signal from spoofing signal, ultimately enabling the generator to produce high-quality synthetic data. Utilizing this strategy, we simulate various typical GNSS spoofing interference scenarios and augment data via GAN across different scenarios.
2. Improved Data Augmentation Using WGAN. To further enhance model detection performance and robustness in complex environments, we designed a WGAN model suitable for GNSS signal data augmentation. The WGAN uses Wasserstein distance to measure the difference between the distributions of real and generated data, ensuring the Lipschitz continuity of the discriminator through gradient penalties. This combination significantly enhances the diversity and quality of generated samples, providing a new perspective for GNSS spoofing interference detection.
Preliminary experimental results demonstrate that GAN can effectively improve detection rates by approximately 4% compared to training sets derived from original datasets when processed through Convolutional Neural Networks (CNN). The WGAN method further increases the detection rate by an additional 1%, achieving an average test accuracy improvement of about 8.5% relative to the original dataset's training set, surpassing the 4.3% improvement provided by GAN. Additionally, WGAN outperforms GAN in reducing false negatives and false positives across all scenarios, generating ten times the amount of data per session compared to GAN. This indicates that our proposed WGAN method has stronger data augmentation capabilities for GNSS data under small sample conditions than traditional GAN methods.