Metric-Based Few-Shot Learning With Triplet Selection for Adaptive GNSS Interference Classification

Abstract

Jamming devices pose a significant threat as they disrupt signals from the global navigation satellite system (GNSS) and thus compromise the accuracy and robustness of positioning systems. The detection of anomalies in frequency snapshots is essential to effectively counteract these interferences. Furthermore, the ability to adapt to diverse and previously unseen interference characteristics is critical to ensuring the reliability of GNSS in real-world applications. In this article, we propose a few-shot learning (FSL) approach to adapt to new classes of interference. We employ pairwise learning techniques, including triplet and quadruplet loss functions, during the training process to enhance the latent representation. In addition, we conducted a benchmark evaluation of state-of-the-art triplet learning methodologies utilizing GNSS datasets. Our method incorporates quadruplet selection, allowing the model to learn representations from various classes of positive and negative interference. Moreover, our quadruplet variant selects pairs based on aleatoric and epistemic uncertainty, facilitating differentiation between similar classes. We evaluated all methods using a publicly available indoor GNSS dataset collected in controlled environments characterized by various multipath effects, and using a dataset obtained from a highway bridge spanning a real-world German highway. Furthermore, we record and publish a second dataset from a highway featuring eight interference classes, in which our FSL method utilizing quadruplet loss demonstrates superior performance in jammer classification accuracy, achieving a rate of 97.66%.