Classification of Spoofed and Non-Spoofed Cross-Ambiguity Function Image by Deep Learning Approach

In this study, transfer learning is used to detect the presence of spoofed signals at the acquisition stage of GNSS signal processing. Transfer learning is a deep learning technique that leverages the pre-knowledge of a pre-trained model from different type of task and data. To do this, an open-source deep learning library, called fast.ai, is used, and the selected model is ResNet-18. It is fine-tuned and evaluated with cross-ambiguity function (CAF) images of clean and spoofed signals in static and dynamic platform from TEXBAT and OAKBAT data. CAF images were generated from the 2D cross-correlation matrix output of each satellite at acquisition using an open-source GNSS software receiver called FGI-GSRx. The model was able to get an average of 94.6% accuracy.