Patrick Xu, Curtis Hay, Rakesh Kumar, Iqbal Surti, Chandra Tjhai; General Motors

View Abstract Sign in for premium content

Abstract:

Most Level 2 autonomous driving systems including General Motors (GM)’ Advanced Driver Assistance System (ADAS) employ multiple sensors to achieve lane-level localization. As the only source onboard that provides absolute 3D global poses, Global Navigation Satellite Systems (GNSS) is susceptible to various Radio Frequency (RF) interferences such as jamming, spoofing and meaconing, intentionally or unintentionally. Since these interferences pose a growing threat to the safety and reliability of the automated vehicles, detecting, characterizing, and mitigating them become increasingly important. This paper proposes a machine learning-based method to detect spoofing signals. Basically, it treats the detection as a binary classification and evaluates the potential of applying supervised machine learning in the process. Various classifiers are trained and evaluated, and the best performing model is further optimized to predict the existence of interference signals. Raw GNSS measurements such as Pseudorange, Carrier Phase, Doppler, Clock Bias and Drift, Carrier-to-Noise Density (CN0), as well as their respective measurement uncertainties, are combined to train and test the machine learning models. Data cleaning, feature engineering, variable correlation analysis, and principal component analysis (PCA) were performed before the training process. Finally, to evaluate the effectiveness of the proposed method, two additional rounds of drive tests with real meaconing signals re-transmitted by multiple GNSS repeaters were performed at a GM vehicle testing facility in Michigan, US. Validation results showed an average of 97.4% detection accuracy, as well as an overall F1 score of 0.937.