RadioML Dataset

The dataset encompasses a wide range of analog and digital modulation schemes, providing a comprehensive training corpus for robust classifiers.

• Analog modulations: AM-DSB, AM-SSB, FM, GMSK
• Digital PSK family: BPSK, QPSK, 8-PSK, OQPSK
• Digital QAM family: 16-QAM, 64-QAM, 128-QAM, 256-QAM
• Total classes: 11 distinct modulation formats in the standard 2016.10A variant, expanding to 24 in the 2018.01A release.

Signals are generated across a wide, discrete SNR range from -20 dB to +30 dB in 2 dB increments. This deliberate staggering allows researchers to evaluate classifier performance from the noise floor to near-perfect channel conditions.

• Low SNR regime (-20 dB to 0 dB): Tests the classifier's ability to operate below the noise floor, a critical requirement for spectrum sensing and SIGINT.
• High SNR regime (+10 dB to +30 dB): Validates baseline performance and the model's ability to handle high-order QAM constellations without confusion.

RadioML 2016.10A provides purely synthetic data with simulated channel impairments, while RadioML 2018.01A introduces over-the-air captures. This dual approach is essential for studying the sim-to-real gap.

• Synthetic channel effects: Includes additive white Gaussian noise (AWGN), multipath Rayleigh fading, sample rate offset, and center frequency offset.
• OTA captures: Generated using USRP software-defined radios in a shielded chamber, introducing real-world hardware impairments like I/Q imbalance and phase noise that are difficult to simulate perfectly.

The dataset stores signals as vectors of complex-valued In-Phase and Quadrature (I/Q) samples, the native format for software-defined radio processing. Each example is a 2x128 vector (I and Q components, 128 time-domain samples).

• Complex-valued representation: Preserves phase information critical for distinguishing PSK and QAM variants.
• Fixed dimensionality: The 128-sample length provides a standardized input tensor for convolutional neural networks (CNNs) and Transformer-based AMC architectures without requiring variable-length sequence handling.

The dataset's labeling structure supports both fine-grained and coarse-grained classification tasks, enabling hierarchical AMC research.

• Fine-grained labels: Directly identify the specific modulation and order (e.g., 16-QAM vs. 64-QAM).
• Coarse-grained families: Group signals by modulation family (e.g., PSK vs. QAM vs. FSK) for hierarchical classification strategies.
• Open-set extension: Researchers can deliberately hold out specific modulation classes during training to benchmark open-set recognition and out-of-distribution detection algorithms.

RadioML has become the de facto standard for reproducible AMC research, with a consistent train/test split and evaluation protocol.

• Standard split: Typically 50% training, 50% testing, stratified across SNR levels and modulation classes.
• Top-1 accuracy metric: The primary evaluation metric is classification accuracy at each SNR level, often plotted as an SNR-vs-accuracy curve.
• Baseline models: The original paper established baselines using a shallow CNN and a deep residual network (ResNet), providing a clear performance floor for subsequent research on Transformer-based AMC and complex-valued neural networks.

A collection of synthetic and simulated radio signals generated using the GNU Radio software-defined radio framework. Unlike RadioML's over-the-air captures, these datasets provide perfect ground truth labels and controlled channel impairments, making them ideal for prototyping AMC algorithms before real-world deployment. Researchers often use GNU Radio datasets to isolate specific signal distortions and validate theoretical classifier performance.

The application of deep neural networks—including CNNs, LSTMs, and Transformers—to learn hierarchical features directly from raw I/Q samples. This approach eliminates the need for hand-crafted feature engineering and has demonstrated superior robustness to channel impairments. Key architectures include:

• Residual Networks (ResNets) for vanishing gradient mitigation
• Transformer encoders for capturing long-range temporal dependencies
• Complex-valued networks that preserve phase information natively

Higher-order statistics (HOS) of a signal's probability distribution that are theoretically immune to additive white Gaussian noise. These hand-crafted features serve as a robust baseline for modulation classification and are often used in hybrid AMC systems that combine traditional feature extraction with neural network classifiers. Second, fourth, and sixth-order cumulants can uniquely identify modulation families like PSK, QAM, and FSK without requiring prior synchronization.

Techniques applied to training I/Q samples to improve model generalization and robustness to real-world channel impairments. Common augmentations include:

• Additive white Gaussian noise injection at varying SNR levels
• Random phase rotation to simulate carrier frequency offset
• Rayleigh and Rician fading emulation for multipath environments
• Time stretching and frequency shifting for Doppler effects These methods are critical for bridging the sim-to-real gap when training on synthetic data.

A classification paradigm where the model must not only identify known modulation schemes but also detect and reject unknown types not seen during training. This is critical for electronic warfare and spectrum monitoring applications where adversaries may deploy novel waveforms. Techniques include:

• Extreme value theory for modeling class boundaries
• Distance-based rejection in learned embedding spaces
• Generative models that synthesize unknown-class examples for training

A methodology where a neural network pre-trained on a large-scale synthetic dataset (like RadioML) is fine-tuned with a small amount of over-the-air data to adapt to specific hardware or channel environments. This approach dramatically reduces the need for expensive real-world data collection while maintaining high accuracy. Domain adaptation techniques further align feature distributions between source and target domains to combat domain shift.