Machine learning-based approach to GPS antijamming

A challenging and outstanding problem in applications that involve or rely on GPS signals is to mitigate jamming. We develop a machine learning-based antijamming framework for GPS signals. Three types of jamming signals are considered: continuous wave interference, chirp and pulse jamming. In addition, white Gaussian noise is assumed to be present. From the point of view of communication, information is encoded in the coarse/acquisition (C/A) code. Multiplying the jammed signal by a sinusoidal wave and integrating over one C/A code period leads to a jammed C/A code signal. To mitigate jamming, we study three types of machine learning methods: reservoir computing (echo state network), multi-layer perceptron, and long short-term memory networks (RNNs). A machine can be trained to learn and predict the signal directly or learn and predict jamming where the real signal can be obtained by removing the jammed component from the total received signal. For a high-frequency carrier (e.g., the standard 1575.42 MHz L1 carrier), learning and prediction can be made computationally efficiently on the C/A code signal. The main result is that machine learning can be effective for predicting and extracting weak GPS signals even in a strongly jammed/noisy environment where the jamming amplitude is three orders of magnitude stronger than the GPS signal. We find that the reservoir computing scheme is stable and performs well for all three types of jamming. The multi-layer perceptron is better for predicting the jamming signal than the GPS signal itself, and the long short-term memory networks work well but only for certain jamming types. In particular, with the direct signal prediction method, the bit error rate (BER) associated with reservoir computing (RC) remains at near-zero values (less than 1%) even for jamming signal ratio (JSR) up to 60 dB for the three types of jamming. The multi-layer perceptron (MLP) method breaks down when the JSR is larger than 20 dB for continuous wave interference (CWI) and pulse jamming, 45 dB for chirp jamming. The long short-term memory (LSTM) can perform very well for the chirp jamming with a near zero error rate and give BER larger than 10% when the JSR is around 40 dB for the CWI and pulse jamming. For the jamming prediction method (indirect method), these three machine learning methods perform well, with near-zero BER (less than 1%). Overall, the RC scheme is stable and performs well for three types of jamming. Besides, RC is fast compared to LSTM method, with much less running time.