Core Concepts
A few-shot learning approach with uncertainty-based quadruplet selection enhances the robustness of interference classification in GNSS data, enabling rapid adaptation to new jammer types.
Abstract
The paper proposes a few-shot learning (FSL) method to adapt a machine learning model for interference classification in global navigation satellite system (GNSS) data. The key contributions are:
Development of a GNSS dataset recorded at a motorway, containing 11 classes - 3 classes of no interference with varying background intensity, and 8 classes of different interference types.
Benchmarking of various FSL techniques, including prototypical networks, matching networks, and relation networks. The prototypical network with Euclidean distance achieves the highest F2-score.
Proposal of a quadruplet loss function that leverages both aleatoric and epistemic uncertainty to select challenging sample pairs for training. This enhances the learned feature representation, leading to improved classification performance compared to the baseline prototypical network and triplet loss.
Evaluation of different adaptation layers, such as convolutional, LSTM, and attention, for post-training the base ResNet18 model. The convolutional layer exhibits the most robust generalization across different class permutations.
The quadruplet loss-based FSL method achieves an accuracy of 97.66% and an F2-score of 0.431 on the GNSS dataset, outperforming the baseline prototypical network (0.363 F2-score) and triplet loss (0.424 F2-score). The learned feature embeddings also show a more continuous representation between similar interference classes.
Stats
The dataset contains 11 classes, with 3 classes of no interference and 8 classes of different interference types.
The dataset is highly imbalanced, with the positive (interference) classes being significantly underrepresented compared to the negative (no interference) classes.
Quotes
"Jamming devices pose a significant threat by disrupting signals from the global navigation satellite system (GNSS), compromising the robustness of accurate positioning."
"The ability to adapt to diverse, unseen interference characteristics is essential for ensuring the reliability of GNSS in real-world applications."