SMART: Spectrum-data for Machine Learning and Analysis through Robust Feature Transformations

University Of Texas At Austin, Austin TX

Investigators

Abstract

This project, called SMART (Spectrum-data for Machine Learning and Analysis through Robust Feature Transformations), aims to enhance access to radio frequency (RF) spectrum data through principled investigation in novel machine learning (ML) architectures that extract characteristic features from raw data. The project promotes the efficient use of spectrum resources critical for communications, national defense, and technological advancement, especially considering the growing interest in spectrum sharing between federal and commercial sectors. By providing a means to collaborate on spectrum data without compromising sensitive information, SMART opens new research and educational opportunities and trust within the ML researchers both in industry and academia. At a broader level, this democratization of data supports the goal of advancing national health and prosperity through improved communication technologies and informed policy decisions. By utilizing latent embeddings, i.e., higher-dimensional features generated through neural networks, instead of raw IQ (in-phase and quadrature) samples, the SMART project addresses significant challenges in spectrum sharing. The key project goals include reducing data storage requirements, facilitating opportunistic data collection, and ensuring the sharing of useful information without compromising sensitive data. SMART is composed of three interconnected research thrusts. The first thrust focuses on Offline Spectrum Feature Engineering, developing methodologies for generating and validating high-dimensional embeddings to support model training. The second thrust, Latency- critical Out-of-Distribution (OOD) Detection and Optimized Labeling, aims to create rapid detection systems for adapting ML models in real time without relying on raw data. Finally, the Toolkit for Feature-Model Validation will empower spectrum owners and regulators to verify the reliability of third-party ML models. By leveraging previous research in long-term evolution (LTE)-radar signal detection and integrating new data collections, SMART seeks to establish a robust framework for trusted spectrum sharing that can benefit both research and industry stakeholders. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.