Siamese Neural Network

The core architectural innovation: two identical neural networks process a pair of input signals simultaneously. These subnetworks share the exact same weights and parameters, ensuring that identical inputs produce identical embeddings. This weight tying guarantees the network learns a consistent metric space—if you swap the two inputs, the computed distance remains symmetric. Each subnetwork acts as a feature extractor, transforming raw IQ data into a compact, high-dimensional embedding vector that captures the hardware impairment signature.

After the twin networks generate two embedding vectors, a distance layer computes the similarity between them. Common choices include:

• Euclidean distance: Straight-line distance in embedding space
• Cosine similarity: Angle between vectors, robust to magnitude variations
• Manhattan distance: Sum of absolute differences

The output is a scalar score—a small distance indicates the same emitter, a large distance indicates different devices. This layer has no trainable parameters; the learning happens entirely in the feature extraction subnetworks.

Training uses a contrastive loss that pulls embeddings of genuine signal pairs together while pushing impostor pairs apart beyond a margin. The loss function is:

L = Y * D² + (1-Y) * max(0, m - D)²

Where Y=1 for genuine pairs (same emitter), Y=0 for impostor pairs (different emitters), D is the computed distance, and m is the margin hyperparameter. This formulation ensures that different-device embeddings are separated by at least the margin, creating a well-structured embedding space for verification.

An alternative training paradigm uses triplet loss, which operates on three inputs simultaneously:

• Anchor: A reference signal from a specific emitter
• Positive: Another signal from the same emitter
• Negative: A signal from a different emitter

The loss function enforces: D(anchor, positive) + margin < D(anchor, negative). This explicitly ensures intra-class distances are smaller than inter-class distances by at least the margin, creating tighter clusters and better separation for open-set recognition scenarios where unknown emitters must be rejected.

At inference time, the network operates as a verification system rather than a classifier:

1. Enrollment: A single reference signal from a known emitter is passed through one subnetwork to generate a stored embedding
2. Query: A new signal is processed through the identical subnetwork
3. Comparison: The distance between the stored and query embeddings is computed
4. Decision: If the distance falls below a calibrated threshold, the signal is authenticated

This enables one-shot learning—a new emitter can be enrolled with a single example, critical for detecting clone devices that were never seen during training.

To ensure the learned similarity metric is invariant to channel conditions rather than transmitter identity, the architecture often incorporates domain-adversarial training. A gradient reversal layer inserted before the distance computation forces the feature extractor to learn representations that are discriminative for emitter identity but indistinguishable for channel conditions. This prevents the network from learning spurious correlations with multipath fading or receiver hardware, ensuring the fingerprint is truly based on transmitter-specific impairments like I/Q imbalance and oscillator phase noise.

The foundational loss function for training Siamese networks. It operates on triplets of samples: an anchor, a positive (same emitter), and a negative (different emitter). The loss penalizes the network when the distance between the anchor and positive is not smaller than the distance between the anchor and negative by a specified margin. This forces RF fingerprints from the same device to cluster tightly in the embedding space while pushing impostors away, enabling robust open-set recognition.

A self-supervised pre-training strategy that teaches the Siamese backbone to extract meaningful representations from unlabeled RF data. The network learns by comparing two augmented versions of the same IQ sample (a positive pair) against samples from other signals (negative pairs). This approach is critical for RF fingerprinting where labeled emitter data is scarce, allowing the model to learn robust, channel-invariant features before fine-tuning on a small set of known devices.

The primary operational advantage of a trained Siamese network. Once the similarity metric is learned, the system can identify a never-before-seen emitter from a single reference capture. The new signal is passed through the twin network alongside the stored reference, and their distance score determines a match. This is essential for clone detection and rapid field deployment, eliminating the need to retrain a classifier every time a new legitimate device is added to the network.

The mathematical function applied to the output embeddings to quantify similarity. Common choices include:

• Euclidean Distance: The straight-line distance between two vectors in the embedding space.
• Cosine Similarity: Measures the angle between vectors, making it robust to differences in signal magnitude.
• Mahalanobis Distance: Accounts for the variance along each feature dimension, useful when certain hardware impairments are more variable than others. The selection directly impacts the system's Equal Error Rate (EER).

A technique to enforce channel robustness in the Siamese embedding. A Gradient Reversal Layer is inserted between the feature extractor and a domain classifier that tries to predict the channel conditions (e.g., indoor vs. outdoor). By training the network to fool this classifier, the learned fingerprint representation becomes invariant to the propagation environment, ensuring the similarity score reflects the transmitter's identity, not the room it was recorded in.

An extension of the Siamese concept for few-shot classification. Instead of comparing pairs, the network computes a prototype vector for each known emitter by averaging the embeddings of a few support examples. A new query signal is then classified by finding the nearest prototype. This is highly effective for emitter identification when only 3-5 reference captures per device are available, a common scenario in signals intelligence.