GNU Radio Conference 2024

TorchSig: A GNU Radio Block & New Tools for Augmenting ML Training with Real World Data

20 Sept 2024, 11:30

15m

Ballroom AB (Knoxville Convention Center (KCC))

Paper (with talk) Machine Learning Main Track

Speakers

Matt Carrick Phil Vallance

Description

The 15-minute talk and paper will consist of three main points: background and description of TorchSig, recent updates and improvements to the system, and future plans for the release of tools to augment training methods using real world data. The goal is to demonstrate a Torchsig ML inference model running within GNU Radio and to also present new machine-learning (ML) tools to the GNU Radio community for training new ML models, performing ML-based signal detection and modulation recognition, and to solicit feedback and ideas for how that is useful to the GNU Radio community.

TorchSig is an open source machine-learning (ML) framework for doing signal detection and modulation recognition. It is able to do modulation recognition over 53 different signals including quadrature amplitude modulation (QAM), frequency shift keying (FSK), Gaussian minimum shift keying (GMSK), different OFDM variants and many others. TorchSig has an extensive library of modulators to generate synthetic waveforms and datasets and utilities which demonstrate how to train an ML model which can then be used for practical applications.

A series of improvements have been made to the codebase, such as reducing the time required to generate datasets, improved DSP implementations such as sidelobe reduction in resampling filters to reduce aliased images, better implementation of G/FSK and G/MSK modulators, and increasing the randomness in datasets to improve training efficiency. Additional DSP and system-level improvements are planned and will be discussed.

A new tool is being developed and will be released to allow ML training to be augmented by real world data using the Label Studio tool. TorchSig currently trains on synthetic modulated IQ data which has been impaired through various transforms, however better ML performance can be obtained by including real world data into the training pipeline. Label Studio allows portions of real world IQ captures to be annotated with modulation labels, allowing the real world data to be used in ML training. The objective is the release of this tool to allow users to train an ML model with their custom datasets. The talk and paper will include examples of this training tool with the real world data from IARPA’s SCISRS.