AeGANet: Adversarially Enhanced Deep Learning Approach for GNSS Interference Discrimination

Technological advancements in Global Navigation Satellite System (GNSS) have greatly brought substantial benefits to positioning services. However, given the inherent vulnerability of GNSS signals to radio-frequency interference (RFI), addressing safety concerns of GNSS-based positioning should be a necessary focus. To support the accurate and effective interference mitigation strategies, this research focuses on the interference nature discrimination, including both interference type and interference power level. Inspired by the profound capabilities of deep learning in image recognition, an improved adversarially-enhanced solution based on Generative Adversarial Network (AeGANet) is proposed, utilizing spectrogram of signals as feeding features. We focus on the pre-correction stage of the receiver, analyzing Intermediate Frequency (IF) signals using Short-Time Fourier Transform (STFT), and feeding the 3-channel spectrogram with power heat mapping into the model. Key innovations of the adversarial model include the label encoding of the generator, the dual-part discriminator design, the detached-objective training approach, as well as the loss function design. These improvements enable the adversarial model to generate and extend the interference dataset, enhancing discriminator training and boosting interference discrimination performance. Evaluated based on a GNSS positioning scenario and an interference test platform with seven signal types across varying interference power levels, the proposed solution demonstrates superior accuracy and performance, laying a solid foundation for the future implementation of effective interference mitigation techniques.