Siamese RF Networks

The defining architectural constraint of Siamese networks: two identical subnetworks process two distinct input samples (e.g., an anchor IQ sequence and a query IQ sequence) with exactly the same weights. This weight-tying guarantees that both inputs are mapped to the same embedding space, ensuring that the computed distance between them is semantically meaningful. Without weight sharing, the network could map identical signals to arbitrary, incomparable locations. During backpropagation, gradients from both branches are summed to update the shared parameters, effectively doubling the batch diversity for the encoder.

Siamese RF networks are trained using metric learning objectives that minimize intra-class distance and maximize inter-class distance. The classic Contrastive Loss takes a pair and a binary label (same emitter or different) and penalizes similar pairs that are far apart and dissimilar pairs that are too close within a margin. Modern implementations often use Triplet Loss, which operates on an anchor, a positive sample, and a negative sample, enforcing that the anchor is closer to the positive than the negative by a fixed margin. This directly optimizes for a k-NN classifier downstream.

The primary operational advantage of Siamese architectures is one-shot learning. After training, the network can identify a new, previously unseen emitter given only a single reference sample. The process works by:

• Encoding the single support sample of the new emitter into an embedding vector.
• Encoding incoming query signals into the same space.
• Classifying the query by finding the nearest neighbor support embedding using Euclidean or cosine distance. This eliminates the need to retrain a classifier every time a new device appears on the network, which is critical for dynamic spectrum environments.

The output of a trained Siamese encoder is a dense, low-dimensional vector (e.g., 128-d or 256-d) that represents the RF fingerprint of a transmitter. These embeddings can be projected into 2D using t-SNE or UMAP for analysis. A well-trained model produces tight, isolated clusters for each unique emitter, even if those emitters were never seen during training. The distance between cluster centroids corresponds to the dissimilarity in hardware impairments. This latent space serves as a semantic map of the RF environment, where proximity implies physical similarity.

The choice of distance function in the embedding space critically impacts verification performance. Common metrics include:

• Euclidean Distance (L2): Sensitive to absolute vector magnitude, standard for triplet loss.
• Cosine Similarity: Measures angular separation, robust to variations in signal power or embedding norm.
• Mahalanobis Distance: Accounts for the covariance structure of the embedding distribution, useful when certain feature dimensions are more variable than others. The distance threshold for accepting or rejecting a match is typically calibrated on a validation set to achieve a target false acceptance rate (FAR).

Training a Siamese network to convergence requires careful selection of training pairs. Hard negative mining is the strategy of selecting negative samples that are difficult to distinguish from the anchor—those that the current model maps close to the anchor in embedding space. Presenting these challenging cases accelerates learning and prevents the loss from flatlining at zero. In RF domains, hard negatives are often signals from different transmitters of the same make and model, forcing the network to learn subtle hardware-specific impairments rather than gross features like bandwidth or center frequency.

The original loss function designed to train Siamese Networks. It operates on pairs of samples, minimizing the distance between embeddings of genuine pairs (same emitter) and maximizing the distance for impostor pairs (different emitters) up to a defined margin. The mathematical objective is L = Y*D² + (1-Y)*max(0, margin - D)², where Y is the binary label (1 for genuine, 0 for impostor) and D is the Euclidean distance between the twin outputs. This loss directly enforces a metric space where emitter identity is defined by proximity.

An advanced ranking loss that uses triplets of samples: an Anchor (A), a Positive (P) from the same class, and a Negative (N) from a different class. The network learns to satisfy the constraint d(A, P) + margin < d(A, N). In RF fingerprinting, this forces the model to pull embeddings of the same transmitter closer together while pushing a different transmitter's embedding farther away than the anchor by a specific margin. Triplet loss is often preferred over contrastive loss because it considers relative distances rather than absolute thresholds, leading to more discriminative RF emitter clusters.

The primary use case for Siamese RF Networks. The model is trained to verify emitter identity from only a single reference sample (the support set). During inference, a new query signal is passed through the twin network alongside the stored reference. If the similarity score exceeds a calibrated threshold, the emitter is verified. This capability is vital for specific emitter identification (SEI) in electronic warfare, where a new radar might only be observed once, and re-identification must occur instantly without retraining the entire classifier.

The defining structural constraint of a Siamese Network. Both branches of the twin network are identical clones, sharing the exact same weight matrices and biases. This symmetry guarantees that two identical input signals will be mapped to the exact same point in the embedding space, regardless of which branch processes them. Weight tying is enforced during backpropagation by summing or averaging the gradients from both branches before updating the single set of parameters. This contrasts with Pseudo-Siamese architectures, which allow branches to have different structures for heterogeneous input types.

The final computational stage fused to the twin outputs. While the Siamese branches produce embeddings, the distance layer computes the similarity score. The standard choice is the L1 (Manhattan) or L2 (Euclidean) distance between the two feature vectors. However, modern implementations often replace simple distance functions with a trainable relation network—a small feed-forward module that learns a non-linear similarity metric from the concatenated or absolute difference of the embeddings. This allows the network to learn complex, domain-specific RF similarity functions beyond simple geometric distance.

The data loading paradigm required for Siamese training. The model is not fed single samples but balanced pairs (genuine/impostor) or triplets. A robust training loop must implement a pair generator that creates these combinations on-the-fly. To prevent the network from taking shortcuts, the generator must maintain a strict 50/50 class balance and implement hard negative mining—selecting impostor pairs that are currently difficult for the network to distinguish. Without this strategy, the network is flooded with easy negatives and fails to learn a tight decision boundary for RF emitter verification.