TorchSig: A GNU Radio Block and New Spectrogram Tools for Augmenting ML Training

Abstract

TorchSig uses machine learning (ML) to detect
and classify digitized radio frequency (RF)
signals. Recent updates and improvements to
TorchSig are given, as well as novel features
for image-only spectrogram generation and training,
reducing memory and computational burdens
and making training much faster. A new
GNU Radio out-of-tree (OOT) block is provided
which uses a TorchSig ML model for detecting
signals in real-time.