Siamese Network

A Siamese network consists of two identical subnetworks that share the exact same weights and configuration. Both branches process their respective inputs—such as two IQ sample bursts—in parallel, producing feature embeddings in a shared latent space. Because the weights are tied, the network is forced to learn a consistent transformation function. This symmetry guarantees that if two signals are from the same device, their embeddings will be geometrically close, regardless of content variation. The architecture is fundamentally a distance metric learner, not a classifier.

The original training objective for Siamese networks is contrastive loss, which operates on pairs of samples labeled as genuine (same device) or impostor (different devices). The loss function:

• Minimizes the Euclidean distance between embeddings of genuine pairs
• Maximizes the distance between impostor pairs beyond a defined margin This margin prevents the network from collapsing all embeddings to a single point. The result is a well-separated embedding space where device identity clusters are distinct and a simple distance threshold can authenticate a transmitter.

A more advanced training paradigm uses triplet loss, which considers three samples simultaneously:

• Anchor: A reference signal from a specific device
• Positive: Another signal from the same device
• Negative: A signal from a different device The loss forces the anchor-positive distance to be smaller than the anchor-negative distance by a specified margin. This relative distance constraint produces more robust embeddings than contrastive loss, as it directly optimizes for ranking rather than absolute distances. Triplet mining strategies are critical for selecting informative triplets during training.

Siamese networks excel at few-shot and one-shot learning scenarios. After training on a diverse set of transmitters, the network can authenticate a newly enrolled device using only a single reference sample. The enrollment process:

• Passes the reference signal through the trained network to compute its embedding
• Stores this embedding as the device's identity vector
• At inference time, compares new signals against this stored vector using a distance metric This eliminates the need to retrain the entire model when adding new devices to the network.

A critical challenge in RF fingerprinting is channel variability—multipath, fading, and environmental changes can distort signal features. Siamese networks address this through domain-invariant feature learning:

• Training pairs are constructed with signals captured under diverse channel conditions
• The shared-weight architecture forces the network to ignore channel-specific artifacts
• The network learns to focus on hardware-intrinsic impairments that persist across environments When combined with data augmentation using synthetic channel models, the embeddings become resilient to real-world propagation effects.

Unlike traditional classifiers with a fixed number of output classes, Siamese networks naturally support open set recognition. The network does not predict a device label; it outputs a similarity score between two inputs. This enables:

• Known device verification: Comparing against enrolled embeddings
• Unknown device rejection: If a signal is not sufficiently similar to any enrolled device, it is flagged as rogue
• Impostor detection: Identifying spoofed or cloned transmitters that fail the similarity threshold This architecture is foundational for zero-trust physical layer authentication systems.

A metric learning loss function that extends the Siamese concept to three inputs: an anchor, a positive sample (same class), and a negative sample (different class). The objective is to minimize the distance between the anchor and the positive while maximizing the distance to the negative by a defined margin.

• Directly optimizes for relative similarity rather than absolute classification.
• Produces highly discriminative feature embeddings for open-set recognition.
• More stable training than contrastive loss for many signal identification tasks.

The process of mapping high-dimensional IQ data or spectrograms into a compact, lower-dimensional vector space. A well-trained Siamese Network produces embeddings where Euclidean distance directly corresponds to transmitter similarity.

• Enables fast nearest-neighbor lookup for device authentication.
• Visualized using t-SNE or UMAP to inspect emitter clusters.
• Serves as the foundation for open set recognition of unknown devices.

Siamese Networks are the canonical architecture for few-shot learning, where a model must authenticate a device after seeing only one or a few enrollment examples. This is critical for rapid IoT onboarding.

• The network learns a similarity function rather than memorizing classes.
• Eliminates the need for retraining when a new transmitter is added to the network.
• Directly applicable to few-shot device enrollment scenarios in secure communications.

A technique to mitigate the distribution shift between training and operational environments. For Siamese fingerprinting models, this ensures the similarity metric remains valid when channel conditions, receiver hardware, or noise floors change.

• Adversarial domain adaptation can align feature distributions across environments.
• Prevents catastrophic degradation in channel-robust feature learning.
• Essential for transitioning models from synthetic digital twin training to real-world deployment.

A classification methodology where the system must identify known emitters while reliably detecting and rejecting unknown or rogue transmitters. Siamese embeddings enable this by thresholding similarity scores.

• Relies on Extreme Value Theory (EVT) to calibrate rejection boundaries.
• The OpenMax algorithm replaces SoftMax for open-set probability estimation.
• Critical for adversarial device spoofing detection in zero-trust wireless architectures.