Convolutional Neural Network (CNN)

The core operation applies learnable filters (kernels) that slide across the input to detect local patterns. In RF fingerprinting, these kernels automatically learn to detect I/Q imbalance artifacts, phase noise skirts, and power amplifier non-linearity patterns without manual feature engineering. Early layers capture simple edges in spectrograms, while deeper layers combine these into complex, device-specific signatures.

CNNs for emitter identification typically process one of three input formats:

• Raw I/Q samples: Complex-valued time-domain data preserving all signal information
• Spectrograms: Time-frequency representations revealing transient and steady-state features
• Bispectrum plots: Higher-order spectral representations resistant to Gaussian noise Each representation exposes different hardware impairment signatures to the convolutional filters.

Pooling layers (max or average) downsample feature maps, providing translation invariance—the ability to recognize a fingerprint regardless of its exact position in the input window. This is critical for RF signals where the precise timing of transient events may shift. Global average pooling often replaces fully connected layers in modern architectures, reducing parameters and preventing overfitting to specific time offsets.

For few-shot device enrollment, CNNs are often deployed in Siamese configurations. Two identical CNN branches process a probe signal and an enrolled baseline, learning a similarity metric in an embedding space. This enables authentication of new devices with minimal training examples by measuring the Euclidean distance between their feature vectors, bypassing the need for retraining the entire network.

Modern emitter identification CNNs employ residual connections (skip connections) that allow gradients to flow directly through deep networks, enabling architectures like ResNet-50 to be adapted for RF data. These deep residual networks capture extremely subtle, high-order interactions between hardware impairments that shallow networks miss, achieving state-of-the-art accuracy in open set recognition tasks where unknown emitters must be rejected.

To prevent the CNN from learning channel-specific artifacts rather than device-specific fingerprints, architectures incorporate domain adversarial training and contrastive learning objectives. These techniques force the network to learn representations that are invariant to multipath fading and channel conditions while remaining discriminative for device identity, ensuring the model generalizes from laboratory training to real-world deployment environments.

The critical pre-processing step that transforms raw I/Q time-series data into a 2D image-like format suitable for a CNN. Techniques like the Short-Time Fourier Transform (STFT) or Continuous Wavelet Transform (CWT) generate spectrograms where hardware impairments manifest as distinct visual textures. This joint time-frequency analysis reveals both steady-state spectral features and transient events, providing the CNN with a rich, spatially organized feature space to learn hierarchical representations of emitter-specific signatures.

The process of using a trained CNN as a deep feature extractor rather than an end-to-end classifier. By removing the final classification layer, the network's penultimate layer outputs a compact, high-dimensional embedding vector that numerically encodes the unique RF fingerprint. This vector serves as the device's identity in a mathematical space where similar emitters cluster together, enabling tasks like authentication via cosine similarity or clustering for open set recognition.

A twin-branch CNN architecture designed for few-shot device enrollment, where only a handful of transmissions are available for training. Instead of learning to classify specific emitters, the network learns a similarity metric. It processes pairs of signal samples through identical, weight-shared convolutional branches and outputs a distance score. This allows a new device to be authenticated by comparing its live signal directly against a stored baseline embedding, without retraining the entire model.

A transfer learning technique that combats channel variability, a primary challenge in real-world RF fingerprinting. A CNN trained on signals from one environment (e.g., a lab) may fail when deployed in another with different multipath characteristics. Domain adaptation methods, such as adversarial training or maximum mean discrepancy minimization, force the network to learn channel-invariant features that focus on the transmitter's hardware impairments rather than the propagation environment.

The high-dimensional vector space where a CNN maps extracted signal features. In this space, the distance between vectors corresponds to device similarity. Key properties include:

• Intra-class compactness: Transmissions from the same device cluster tightly together.
• Inter-class separability: Clusters for different devices are far apart.
• Cosine similarity or Euclidean distance is used to verify identity by measuring the proximity of a probe signal's embedding to an enrolled baseline.

A classification paradigm where the CNN must identify known emitters while simultaneously detecting and rejecting any transmitter not present in the training database. This is achieved by analyzing the embedding space and setting a distance threshold. If a new signal's embedding falls far from all known clusters, it is flagged as an unknown or rogue device. This is critical for spectrum surveillance and intrusion detection in dynamic electromagnetic environments.