Return to Session D6 Next Abstract

Session D6: AI/Machine Learning (ML) for PNT Situational Awareness

GPS Situational Awareness Using Unsupervised Machine Learning Techniques
Stanley Radzevicius, Brian Sexton, Wilbur Myrick, ENSCO, Inc.
Location: Ballroom C
Date/Time: Wednesday, Jun. 14, 10:50 a.m.

It is difficult to monitor and characterize a contested NAVWAR GNSS spectrum at any given point in time. With sophisticated agile threats, it is important to monitor the NAVWAR GNSS spectrum without the need to have any prior knowledge of the type of threats. ENSCO, Inc. is developing unsupervised machine learning methods designed to detect malicious RF activity in the GNSS spectrum. These methods can be used to detect anomalous and malicious activity in GNSS RF bands. This provides situational awareness in the presence of jamming threats and other activity of interest that is not normally present.
In previous efforts, ENSCO has developed supervised deep learning techniques to assess GPS integrity. This involved collection of training data on multiple sensor modalities, development of physical models, and implementation of deep learning methods to produce an integrity score. This architecture is very effective at detecting threats (compromised GPS) if a sufficient amount of labeled training data is available to develop the supervised algorithms. Additional sensor modalities help detect threats in more complex environments. The unsupervised methods discussed here are beneficial when training data are unavailable.
ENSCO plans to present an unsupervised machine learning architecture to identify malicious RF activity in the L band. These machine learning techniques leverage only “trace data” over time from a Software Defined Radio (SDR), not raw I/Q. Anomalies and processes must be observable using only magnitude and frequency (no phase data) to be detected. The unsupervised learning methodologies can incorporate additional features such as phase when new data types become available to extend detection capabilities.
For a unimodal background distribution, normalized cross-correlations combined with automated statistical metrics are effective at detecting outliers and anomalous activity to provide an RF pattern of life (POL) of malicious activity in the NAVWAR GNSS spectrum without ground truth data. For a multimodal background distribution this approach requires identifying the modes in the data and computing normalized cross-correlations for each mode to obtain accurate statistical metrics used to identify outliers and other anomalous activity.
An unsupervised subspace algorithm is used to identify the modes present in the multimodal data and detect outliers in both unimodal and multimodal data. This unsupervised machine learning algorithm is also used to elucidate the processes generating GPS pattern-of-life for integrity monitoring. The unlabeled GPS spectrogram data contain benign background, jamming, and spoofing processes that are a function of time. The high dimensionality of the spectrogram data is reduced for improved clustering and data visualization of clusters that provide insight in the nature and number of processes producing the spectral data. An efficient latent subspace representation is needed to preserve relevant information using only a small number of features in the dimensionality-reduced subspace. Score plots are color-coded to visualize changes in data processes as a function of time. The unsupervised learning algorithm has been successfully applied to Wi-Fi and can be applied to other pattern-of-life applications.
This unsupervised technique can be used to alert a user on the presence of GNSS anomalies when no prior information is known. It provides anomaly detection and differentiation as well as characterization of the RF POL. The subspace clusters when combined with labeled data can serve as a foundation for supervised machine learning algorithms to classify GPS interference.
Data for algorithm development and testing was acquired at the PNTAX 2022 event (to include benign and contested environments) and additional benign data was captured near the ENSCO facility in Springfield, VA. Initial results showed accurate clustering of samples when effects were present in the contested GPS environment. The algorithms accurately displayed one cluster during benign GPS collections. Future work includes incorporating I/Q data or other channels to establish a higher fidelity POL and thus detect anomalies with higher confidence.

Return to Session D6 Next Abstract