Siamese Neural Network for RF

Two identical subnetworks process a reference signal and a test signal in parallel. Both branches share the exact same weights, ensuring that identical signals map to identical points in the embedding space. This symmetry forces the network to learn a canonical feature representation rather than memorizing specific transmitters.

• Each branch typically uses a Convolutional Neural Network (CNN) or Transformer backbone optimized for I/Q samples.
• Weight sharing guarantees that the distance metric learned is consistent and symmetric.
• The architecture naturally handles variable-length signal inputs by using global pooling before the embedding layer.

The network is trained with a contrastive loss that minimizes the Euclidean distance between embeddings of signals from the same transmitter while maximizing the distance between embeddings from different transmitters beyond a specified margin.

• Loss formula: L = Y*D² + (1-Y)*max(0, margin - D)² where Y=1 for genuine pairs and Y=0 for impostor pairs.
• The margin parameter defines the radius of the cluster for each transmitter identity.
• This loss directly optimizes for inter-class separation and intra-class compactness in the embedding space.

An alternative training objective uses triplets of signals: an anchor (reference signal), a positive (same transmitter), and a negative (different transmitter). The loss ensures the anchor-positive distance is smaller than the anchor-negative distance by at least a margin.

• Triplet mining strategies like semi-hard negative mining are critical for stable convergence.
• This formulation is more robust to imbalanced datasets where some transmitters have many more samples than others.
• Often combined with online triplet generation within each training batch to maximize discriminative power.

The output of each branch is a compact embedding vector (typically 128-256 dimensions) that represents the RF fingerprint of the input signal. During inference, a reference embedding is computed once per known device and stored in a gallery database.

• New test signals are identified by nearest-neighbor search in the embedding space using cosine or Euclidean distance.
• This enables one-shot learning: a new transmitter can be enrolled with a single reference capture, without retraining the entire model.
• The embedding space naturally supports open-set recognition by thresholding the similarity score to reject unknown devices.

To prevent the network from learning channel artifacts instead of transmitter hardware signatures, domain adversarial training is often integrated. A gradient reversal layer forces the feature extractor to produce embeddings that are invariant to channel conditions while remaining discriminative for device identity.

• The adversarial branch classifies channel type (e.g., indoor, outdoor, fading profile) and its loss is subtracted from the main objective.
• Data augmentation with simulated channel impairments (multipath, Doppler, noise) during training further improves robustness.
• The result is a fingerprint that captures amplifier non-linearity and I/Q imbalance rather than environmental propagation effects.

The final verification decision relies on a similarity score computed between the two embedding vectors. Common metrics include cosine similarity, Euclidean distance, or a learned metric network that takes both embeddings as input and outputs a similarity probability.

• Cosine similarity is preferred when embedding magnitudes may vary due to signal power differences.
• A learned metric layer (e.g., a small fully-connected network) can capture non-linear decision boundaries in the embedding space.
• The Equal Error Rate (EER) operating point is determined by sweeping the similarity threshold to balance false acceptance and false rejection rates.

Identify a previously unseen, unauthorized transmitter from a single capture by comparing its embedding against a gallery of known authorized devices. The Siamese network computes a similarity score; if the score falls below a calibrated threshold, the device is flagged as a rogue.

• Use Case: Detecting an intruder radio attempting to join a secure tactical mesh network.
• Key Metric: Equal Error Rate (EER) below 0.1% for high-SNR captures.
• Advantage: No retraining required when new authorized devices are added—simply enroll their embeddings.

Re-identify a specific transmitter as it hops across frequencies or switches modulation schemes. The Siamese network learns a channel-robust embedding space where the fingerprint of Device A on FSK at 900 MHz is close to Device A on QPSK at 2.4 GHz, but far from Device B.

• Training Strategy: Contrastive loss pulls same-device pairs together and pushes different-device pairs apart, regardless of channel or modulation.
• Operational Impact: Persistent tracking of a target emitter during frequency-hopping spread spectrum (FHSS) operations.

Distinguish a genuine transmitter from a sophisticated hardware clone that shares the same make, model, and firmware revision. The Siamese network amplifies microscopic, non-cloneable impairments—such as I/Q imbalance and oscillator phase noise—that remain unique to each physical die.

• Method: Embed genuine and suspect signals; measure cosine distance in the learned metric space.
• Result: Reliable separation of clones even when traditional feature engineering (EVM, carrier frequency offset) shows overlap.

Build a dynamic allowlist of transmitters without pre-defining classes. The Siamese backbone acts as a feature extractor; new devices are enrolled by storing their embedding vectors. At inference, a query signal is compared against all enrolled embeddings using the learned similarity function.

• Architecture: Siamese encoder + nearest-neighbor classifier.
• Benefit: Seamlessly add or revoke devices without model retraining, critical for ad-hoc coalition networks.
• Rejection Mechanism: If no enrolled embedding exceeds the similarity threshold, the query is classified as unknown.

Detect replay attacks and adversarial signal manipulations designed to fool traditional classifiers. A Siamese network trained with adversarial contrastive examples learns to ignore low-level perturbations while focusing on immutable hardware signatures.

• Threat Model: An attacker captures and retransmits a legitimate signal with added noise or time shift.
• Defense: The network compares the live signal against the enrolled embedding; replay artifacts and channel inconsistencies reduce the similarity score below the acceptance threshold.

Rapidly adapt an existing Siamese fingerprinting model to a new operational environment—such as a different frequency band or geographic region—using only 5-10 labeled examples per device. The pre-trained embedding space provides a strong prior; fine-tuning with a small support set aligns the manifold to the target domain.

• Technique: Meta-learning with MAML or prototypical network initialization.
• Operational Value: Deploy a working SEI system in a new theater of operations within minutes, not weeks.

The Siamese network's twin branches project raw I/Q samples into a high-dimensional embedding space where Euclidean distance corresponds directly to transmitter similarity. Signals from the same device cluster tightly together, while different devices form well-separated manifolds.

• L2 normalization constrains embeddings to a hypersphere for stable training
• Cosine similarity is often used as the distance metric for inference
• Visualization with t-SNE or UMAP reveals the quality of the learned representation

The geometry of this space determines the system's ability to reject rogue devices while accepting legitimate ones.

A critical challenge for any RF identification system. Multipath fading and environmental changes can distort a signal's fingerprint more than the hardware impairments themselves. Siamese networks can be trained with domain adversarial techniques to learn channel-invariant representations.

• Data augmentation simulates diverse channel conditions during training
• Domain adversarial training forces the feature extractor to confuse a channel classifier
• Contrastive pre-training on unlabeled RF data builds robustness before fine-tuning

Without channel robustness, a model trained in a lab will fail when deployed in the field.