Novelty Detection for RF Waveforms with Ensembled Contrastive Learning
Abstract
We consider the problem of identifying whether an observed radio-frequency (RF) waveform belongs to a known class of waveforms or is unfamiliar. For example, this problem arises in spectrum monitoring of communication or radar signals, where there is a need to detect illegal or unknown transmitters operating in the range of the receiver. When the known waveforms are described by a dataset rather than a mathematical model, our problem is an instance of \novelty detection" from the eld of machine learning. In this dissertation proposal, we describe research on novelty detection for RF waveforms that leverages deep neural networks and recent advances from the eld of contrastive learning. One of our main contributions is a method for ensembling several detection scores, which we call shift-ensembled novelty detection (SEND). Our technique is shown to increase detection performance over existing novelty detectors on communication and radar datasets.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 2022
- Accession Number
- AD1167331
Entities
People
- Edward Reehorst
Organizations
- Air Force Research Laboratory