GNU Radio Dataset

GNU Radio datasets are created using software-defined radio (SDR) flowgraphs that simulate the entire digital communications chain. This includes bit generation, modulation mapping, pulse shaping, and channel impairment simulation. The framework allows precise control over every parameter, enabling the generation of perfectly labeled I/Q samples for supervised learning. Common modulation schemes include BPSK, QPSK, 16-QAM, and 64-QAM, often generated with varying root-raised cosine (RRC) roll-off factors.

A key advantage of synthetic datasets is the ability to systematically introduce and isolate real-world impairments. GNU Radio flowgraphs can add:

• Additive White Gaussian Noise (AWGN) across a defined SNR range
• Multipath fading using models like Rayleigh or Rician
• Carrier Frequency Offset (CFO) and Sample Rate Offset (SRO)
• Phase noise and IQ imbalance This allows researchers to benchmark AMC model robustness against specific hardware and channel non-idealities.

Datasets like the seminal RadioML series (e.g., RadioML 2016.10A, RadioML 2018.01A) were generated using GNU Radio. These open-source benchmarks provide a standardized testbed for comparing deep learning AMC architectures. They typically contain millions of labeled I/Q samples across 11 modulation types and a wide SNR range, enabling reproducible research and accelerating the development of convolutional neural network (CNN) and transformer-based classifiers.

Despite their utility, GNU Radio datasets suffer from a sim-to-real domain gap. Simulated channel models are imperfect approximations of physical electromagnetic environments. Models trained solely on synthetic data often degrade when deployed on over-the-air (OTA) captures due to unmodeled hardware impairments, non-Gaussian interference, and dynamic multipath. This motivates domain adaptation and transfer learning techniques to fine-tune synthetic-trained models with a small amount of real-world data.

Each generated signal is accompanied by rich metadata, including the modulation type, SNR, symbol rate, and pulse shaping filter parameters. This perfect labeling is impossible to obtain with blind OTA captures. The metadata enables granular performance analysis, such as plotting classification accuracy versus SNR to identify the SNR wall for a specific modulation scheme. This structured data is critical for feature-based AMC methods that rely on known statistical properties like cumulants.

GNU Radio's modular block architecture allows researchers to extend standard datasets with custom signal types. This includes generating higher-order QAM (256-QAM, 1024-QAM), OFDM waveforms, or even emulating specific electronic warfare (EW) threats. The framework supports hierarchical AMC research by generating datasets with both intra-class (e.g., QAM order) and inter-class (e.g., PSK vs. QAM) labels, enabling the development of multi-stage classification pipelines.

Techniques that artificially expand training datasets by applying realistic channel impairments to I/Q samples. Common augmentations include:

• Additive White Gaussian Noise (AWGN) at varying SNR levels
• Phase rotation to simulate carrier frequency offset
• Rayleigh and Rician fading to model multipath environments
• Time stretching to simulate symbol rate mismatch These methods improve model generalization to real-world channel conditions without requiring expensive over-the-air data collection.

A neural network architecture that natively processes complex-valued I/Q data using complex weights and activation functions. Unlike real-valued networks that split I/Q into two separate channels, CVNNs preserve the phase information inherent in the complex baseband representation. This enables more compact models that learn rotation-equivariant features, improving classification accuracy under carrier frequency offset and phase rotation impairments.

A methodology where a neural network is pre-trained on a large-scale synthetic dataset (such as GNU Radio-generated signals) and then fine-tuned with a small amount of over-the-air data. This approach addresses the domain shift between simulated and real-world RF environments. The pre-trained model learns generalizable signal features, while fine-tuning adapts the classifier to specific hardware impairments, antenna characteristics, and propagation conditions of the target deployment.

A classification paradigm where the model must not only classify known modulation schemes but also detect and reject unknown modulation types not seen during training. In electronic warfare contexts, GNU Radio datasets typically contain only known threat modulations. Open-set recognition prevents the model from making forced, incorrect classifications when encountering novel waveforms, instead flagging them for human analysis or triggering out-of-distribution detection protocols.

A self-supervised training method that learns robust signal representations without requiring labeled data. The technique works by:

• Pulling augmented views of the same I/Q sample together in embedding space
• Pushing views from different samples apart This pre-training strategy is particularly valuable for GNU Radio datasets, as it can leverage the vast amounts of unlabeled synthetic signals before fine-tuning on a smaller labeled subset for specific modulation recognition tasks.