Previous Abstract Return to Session C8 Next Abstract

Session C8: Integrity and Assurance 2

Machine Learning Based GPS Authentication for PNT Open Systems Architecture
Brian Sexton, Stanley Radzevicius, Wilbur Myrick, ENSCO, Inc.
Location: Ballroom C
Date/Time: Wednesday, Jun. 8, 4:45 p.m.

ENSCO, Inc. is developing a multiclass machine learning framework for PNT situational awareness that supports the Army Assured-Positioning, Navigation, and Timing (A-PNT) Open Architecture and CMOSS compatible platforms. This framework is designed to provide a GPS integrity monitoring solution for general purpose processors leveraging CMOSS compatible Software Defined Radios (SDRs). A prototype software architecture will be discussed that uses a modular, multi-input machine learning (ML) data architecture that integrates mixed data types and optimally blends the individual sensor data ML inference branches to provide PNT situational awareness. This sensor-combining approach supports integration of diverse PNT data and provides multi-input scalability necessary to realize the modularity targeting CMOSS compliant A-PNT cards.
ENSCO plans to present a GPS receiver authentication application leveraging this machine learning framework applicable to CMOSS compliant A-PNT cards. This research focuses on the ability to leverage Commercial-off-the-Shelf (COTS) Software Defined Radios (SDRs) to discriminate between transmitters associated with GPS signals when combined with our Machine learning (ML) framework. ENSCO seeks to develop and integrate a Machine Learning based GPS Authentication algorithm that supports PNT situational awareness leveraging COTS SDRs for GPS authentication. A Machine Learning architecture is discussed that provides GPS transmitter discrimination for a given COTS SDR. The research will focus on the ability to extract features such that a machine learning algorithm can identify one specific transmitter from another.
The multiclass machine learning framework consists of open-source software programmed using Python. TensorFlow and Keras are used to build the deep learning models. To develop a workable RF machine learning authentication algorithm framework for adaptable PNT systems, ENSCO explores features based on the GPS correlator outputs. One-dimensional convolutional neural networks are used to learn salient feature representations directly from the IQ sequences generated by the GPS correlators. The flexible machine learning architecture supports a variable number of sequence inputs and can accommodate disparate data types. The final classification layer of the neural networks consists of a densely connected layer with a softmax activation that assigns probability scores to each of the space vehicle numbers.
ENSCO plans to explore the process of finding reliable machine learning feature differences that support an RF machine learning authentication algorithm based on GPS transmitter classifications. The RF signal differences between two copies of the same transmitter are very small and subtle. It is even more difficult depending on the type of COTS SDRs that are used to generate machine learning features based on the GPS correlator outputs. ML results are presented exploring the discrimination capability of COTS SDRs to support GPS authentication between the various GPS transmitters. This research provides a path for deploying a generalized ML framework for PNT situational awareness and GPS authentication that extends to CMOSS compliant A-PNT cards.



Previous Abstract Return to Session C8 Next Abstract