Previous Abstract Return to Session B3 Next Abstract

Session B3: AI and Machine Learning for GNSS Signal Processing, Error Modeling, and Threat Detection

A Supervised Architecture for GPS Interference Classification and Agile PNT Situational Awareness
Noah Lewis, National Research Council Post-Doctoral Fellowship Program, Air Force Research Laboratory, Space Vehicles Directorate; Erik Rennspiess, Air Force Research Laboratory, Space Vehicles Directorate; Anisa Patel, Georgia Institute of Technology; Brooke Howell, Scott Minas, Khanh Pham, Air Force Research Laboratory, Space Vehicles Directorate
Location: Windsong 7-8
Date/Time: Thursday, Sep. 17, 9:20 a.m.

This work addresses the growing vulnerability of GPS signals within an overcrowded and challenged electromagnetic spectrum. The low-power nature of these signals makes them highly susceptible to both intentional and unintentional jamming as well as false signals presenting as legitimate ones known as spoofing. Current mitigation techniques rely on a protracted, human-in-the-loop process: signal capture, offline analysis by engineers (often taking months or years), waveform library development, and eventual filter deployment. This reactive approach gives adversaries, capable of deploying novel, uncatalogued interference waveforms the advantage.
Our effort is to develop and test a complete, end-to-end architecture that leverages Machine Learning (ML) to radically shorten this cycle. We are developing a system to perform real-time, in-situ recognition, characterization, and automatically sorting signals to known types (classification), with a primary focus on the GPS L-band. The work targets the full spectrum of interferers, from basic continuous-wave or broadband noise jammers to sophisticated, and structured threats like chirp jammers.
The central scientific inquiry driving this research is: Can we engineer a system that fully automates the Observe, Orient, Decide, and Act (OODA) loop for GPS signal processing in NAVWAR environments; creating a cognitive radio capability for PNT assurance. The focus of this work is the first steps of this system, Observe and Orient. This involves investigating if ML models like Convolution Neural Networks (CNNs) or Recurrent Neural networks (RNNs) can extract high-dimensional features from raw I/Q data to not only classify known interferers with high accuracy but also perform anomaly detection to flag and characterdize previously unseen "zero-day" waveforms. The ultimate goal is an autonomous decision engine that maps signal classification to a specific, effective countermeasure with near-zero latency.
At the core of our architecture is an algorithmic methodology that receives GPS signals, cleans the data, analyzes it, and returns important statistics about the interference patterns. Our methodology is a pipeline that leverages spectral information extracted from In-phase/Quadrature (I/Q) data to train an advanced feature reduction method, using contrastive learning, that distills the data into its most basic signal, removing as much noise as possible. This model uses information within the data to compare and contrast differences between possible interference signals to reduce the I/Q data into only interference relevant information. These reduced, information-heavy feature spaces are now effective for possible down-stream tasks including interference classification, interference avoidance, and source localization. This work focuses on the distillation process while showing its effectiveness by developing a novel interference classification algorithm. Interference classification is a key part of aiding the
warfighter and increasing battlefield awareness. By detecting and quantifying the interference pattern, we can build additional algorithmic solutions to for problems such as threat assessment, interference avoidance, and source analysis. This work represents just the first step in a fully realized system that can aid the warfighter with battlefield awareness.

Previous Abstract Return to Session B3 Next Abstract