Lightweight GNSS Interference Classifier and Its Solution Under Few-Shot Conditions

The inherent vulnerabilities of the Global Navigation Satellite System (GNSS) make it highly susceptible to various intentional or unintentional interferences. GNSS interference identification is crucial for the targeted selection of anti-jamming measures. Existing GNSS interference identification methods primarily rely on deep learning techniques, but typically using single-scale convolutional kernels to extract time-frequency features of interference signals, failing to fully exploit the multi-scale information present in time-frequency images. Additionally, existing methods use a Multi-Layer Perceptron (MLP) as the decision layer of the classifier. However, the activation functions of each layer in the MLP are fixed, which somewhat limit the model's fitting ability and flexibility. Finally, existing GNSS interference identification methods usually require a large number of interference signal samples to train the classifier. It is extremely difficult to obtain sufficient interference signal samples in a real electronic warfare environment. Therefore, this paper first proposes a GNSS interference classifier based on Multi-Scale Feature Fusion (MSFF) and Kolmogorov-Arnold Network (KAN), also known as MSFF-KAN. To improve the identification performance under few-shot conditions, this paper further adds a Spatial Transformer (ST) module to increase the number and diversity of training samples. The experimental results show that when there are sufficient training samples, the accuracy of the MSFF-KAN exceeds 95%. Even when there is only 1 training sample for each type of interference, the MSFF-KAN with the ST module has an accuracy improvement of at least 18% compared to existing GNSS interference classifiers.