Also known as: RFML, RF machine learning, RF deep learning, deep learning for RF
RF machine learning (RFML) is the application of machine learning — usually deep neural networks — to radio-frequency signals, learning directly from IQ samples or spectrograms rather than from hand-engineered detectors.1 Instead of a human deriving the exact statistic that distinguishes QPSK from 8PSK, a network is trained on many labelled examples and discovers the discriminating features itself. Core RFML tasks include automatic modulation classification, broader signal classification, interference and anomaly detection, and device RF fingerprinting (identifying a specific transmitter from subtle hardware imperfections).
How it works
An RFML system has two phases. Training happens offline on a big machine: a labelled dataset of recordings is fed to a network — often a 1-D convolutional net over raw IQ, or a 2-D CNN over spectrogram images — which adjusts its weights to minimize classification error. This is where the well-known public datasets come in, such as the DeepSig RadioML sets that pair many modulation types across a range of SNRs.2 Because signals are notoriously variable, training leans heavily on data augmentation and channel simulation: adding noise, frequency and timing offsets, and fading so the model generalizes beyond the lab.
Inference is the deployed phase: a short window of live IQ (or its spectrogram) is passed through the frozen model, which outputs class probabilities. Inference is far cheaper than training and can run on a CPU or a small accelerator, but it inherits the training set’s blind spots — a model only recognizes what it was taught, and confidence scores can be misleadingly high on out-of-distribution signals.
Reproducibility hinges on metadata. A recording is useless for ML without its center frequency, sample rate, and labels, so RFML datasets are increasingly packaged in SigMF, the open standard that stores IQ alongside a JSON metadata and annotation file. Clean, well-labelled, SigMF-style data is often the hardest and most valuable part of an RFML project.
Relevance to SDR
RFML is an active research and increasingly operational field: cognitive radio and dynamic spectrum access, spectrum-monitoring systems that flag unknown emitters, interference hunting, and security work that fingerprints or spoofs devices. Its promise is handling the unknown — signals no hand-written decoder was built for — and adapting as the spectrum changes. Its limits are real too: heavy data and compute needs, brittleness under distribution shift, and poor interpretability compared with a deterministic decoder whose every step is auditable. The training side typically wants a GPU or accelerator; inference is what gets pushed toward the edge.
GopherTrunk does not use machine learning. Its entire chain is classical, deterministic DSP and protocol logic — down-conversion, matched filtering, timing and carrier recovery, symbol slicing, error correction, and framing — and it decodes signals whose modulation and framing are already known (P25, DMR, NXDN, TETRA, and the rest), so there is nothing to classify: the mode is specified, not inferred. That makes GopherTrunk exact and repeatable but, by design, blind to signals it has no decoder for. RFML addresses the complementary problem of recognizing and sorting unknown signals, and the two could sit side by side in a monitoring stack — an RFML classifier triaging the spectrum and handing known modes to a deterministic decoder like GopherTrunk — but GopherTrunk itself ships no models, no training, and no inference.
Sources
-
Machine learning — Wikipedia, for the supervised-learning framework RFML applies to RF signals. ↩
-
Convolutional Radio Modulation Recognition Networks — O’Shea, Corgan, Clancy (2016), the paper behind the RadioML datasets and CNN-on-IQ modulation recognition that seeded modern RFML. ↩