GNSS Interference Classification Using Handcrafted Signal Features and Lightweight Neural Models

Abstract

This paper compares three approaches to six-class Global Navigation Satellite System (GNSS) interference classification: classical classifiers trained on 17 handcrafted spectral and statistical features, a one-dimensional convolutional neural network using two-channel raw IQ samples, and a two-dimensional convolutional neural network using STFT spectrograms. Experiments use the public Raw IQ dataset and its prescribed split of 6,000 training and 1,500 test samples. A random forest and a multilayer perceptron classified all test samples correctly, while logistic regression, the raw-IQ network, and the spectrogram network achieved macro-F1 scores of 0.9993, 0.9980, and 0.9940. The leading methods differ by only a few test samples, so the experiment does not establish a meaningful accuracy advantage among them. The practical differences instead concern model complexity and inference cost. On the same CPU, median model-only latency was 0.141 ms for logistic regression, 7.52 ms for the raw-IQ network, and 35.2 ms for the spectrogram network; logistic regression used only 108 fitted coefficients. These results indicate that handcrafted features provide the most favorable accuracy-complexity trade-off on this synthetic high-interference benchmark, although validation on measured and lower-interference signals remains necessary.