Use of Supervised Machine Learning for GNSS Signal Spoofing Detection with Validation on Real-World Meaconing and Spoofing Data—Part I

Abstract

The vulnerability of the Global Navigation Satellite System (GNSS) open service signals to spoofing and meaconing poses a risk to the users of safety-of-life applications. This risk consists of using manipulated GNSS data for generating a position-velocity-timing solution without the user’s system being aware, resulting in presented hazardous misleading information and signal integrity deterioration without an alarm being triggered. Among the number of proposed spoofing detection and mitigation techniques applied at different stages of the signal processing, we present a method for the cross-correlation monitoring of multiple and statistically significant GNSS observables and measurements that serve as an input for the supervised machine learning detection of potentially spoofed or meaconed GNSS signals. The results of two experiments are presented, in which laboratory-generated spoofing signals are used for training and verification within itself, while two different real-world spoofing and meaconing datasets were used for the validation of the supervised machine learning algorithms for the detection of the GNSS spoofing and meaconing.

Use of Supervised Machine Learning for GNSS Signal Spoofing Detection with Validation on Real-World Meaconing and Spoofing Data—Part II

“…The correlation patterns within those data can be established and monitored with the purpose of distinguishing between false and authentic GNSS measurements and observables. In addition, the application of an SVM classifier to GNSS spoofing detection has been previously demonstrated as promising by [8][9][10][11].…”

“…As was described in depth in Part I of this publication [11], various datasets for the C-Support Vector Machine model building and its validation have been used. As in Experiments I and II [11], all three groups of GNSS datasets are also used in Experiments III and IV.…”

“…In Part I [11], Experiments I and II were conducted by using the synthetically generated spoofing dataset for training dataset formation followed by the C-SVM model building, with eventual validation of the model by the meaconing validation dataset (Experiment I) and spoofing validation dataset (Experiment II).…”

“…Results of applying C-SVM in the Experiments I and II [11] showed that a relatively high number of supporting vectors has been involved in the separation between the authentic and false GNSS observables, which points to the higher complexity of the C-SVM model with potential overfitting. However, in the results of model validation by two independent datasets, the meaconing and spoofing datasets in Experiment I and II, respectively, demonstrated a high success rate and a lack of overfitting.…”

“…However, in the results of model validation by two independent datasets, the meaconing and spoofing datasets in Experiment I and II, respectively, demonstrated a high success rate and a lack of overfitting. Following these findings [11], we wanted to examine the effect of enriching the training data for the supervised machine learning-based approach with the real-world spoofing and meaconing datasets, eventually validating the proposed approach on these independent, real-world spoofing and meaconing datasets. In more details, we wanted to gain a deeper insight into the characteristics of the real-world spoofing and meaconing datasets, and potential implications these characteristics might have within the supervised machine learning-based approach for detection of the GNSS signal manipulation attempts.…”