Spectrogram

A deep learning architecture that applies convolutional filters to spectrogram images, automatically learning spatial hierarchies of features. CNNs excel at detecting localized time-frequency patterns such as transient pulses, carrier drift, and modulation artifacts without manual feature engineering.

• Uses 2D convolutional layers to scan spectrograms for discriminative patterns
• Pooling layers provide translation invariance to minor time and frequency shifts
• Architectures like ResNet and EfficientNet are commonly fine-tuned for RF tasks

A neural architecture that applies self-attention mechanisms to spectrogram patches, capturing long-range dependencies across the entire time-frequency representation simultaneously. Unlike CNNs, transformers model global context without being constrained by local receptive fields.

• Vision Transformer (ViT) treats spectrogram patches as a sequence of tokens
• Excels at identifying wideband spectral signatures spanning large frequency ranges
• Self-attention weights can be visualized to reveal which time-frequency regions drive classification decisions

The process of applying realistic channel transformations to spectrogram images to expand training datasets and improve model robustness. Augmentation simulates operational variability without requiring additional real-world captures.

• Additive white Gaussian noise injection mimics varying signal-to-noise ratios
• Time stretching and frequency shifting simulate Doppler effects and oscillator drift
• SpecAugment randomly masks time or frequency bands to force the model to learn redundant features

A technique where a model pre-trained on a large source dataset of spectrograms is fine-tuned on a smaller target dataset of specific emitters. This dramatically reduces the number of labeled samples required for new device enrollment.

• Models pre-trained on ImageNet can be adapted despite the domain gap from natural images
• Domain-specific pre-training on large-scale RF datasets yields superior performance
• Only the final classification layers are typically retrained, preserving learned feature extractors

The process of mapping spectrogram inputs into a compact, lower-dimensional vector space where signals from the same device cluster tightly together. Embeddings enable efficient similarity comparisons and open set recognition.

• Siamese networks learn embeddings by comparing pairs of spectrograms
• Triplet loss ensures intra-class distances are smaller than inter-class distances
• Visualization via t-SNE or UMAP reveals natural clustering of emitter identities

A classification paradigm where the model not only identifies known emitters but also reliably rejects unknown transmitters not seen during training. This is critical for detecting rogue devices and spoofing attempts in operational environments.

• OpenMax replaces standard SoftMax with a calibrated rejection mechanism based on Extreme Value Theory (EVT)
• Models analyze the distance of embeddings from known class centroids
• Thresholds are calibrated to balance detection sensitivity against false alarm rates