Contrastive Learning

The core mechanism relies on defining positive pairs (signals from the same device) and negative pairs (signals from different devices). In RF fingerprinting, a positive pair might be two different signal bursts from the same transmitter, while a negative pair consists of bursts from two distinct devices. The model learns to minimize distance between positive pairs while maximizing distance between negatives, creating a structured embedding space where device identity clusters naturally emerge.

Contrastive learning creates supervisory signals from the data itself without manual labels. Common pretext tasks for RF signals include:

• Temporal cropping: Treating different time segments of the same transmission as positive pairs
• Augmentation invariance: Applying channel distortions, noise, or frequency shifts and requiring the model to recognize the same underlying device
• Multi-view coding: Using different signal representations (I/Q samples, spectrograms, cyclostationary features) as different views of the same emitter

The InfoNCE (Information Noise-Contrastive Estimation) loss is the dominant objective function. It frames representation learning as a classification problem where the model must identify the true positive sample among a batch of negative distractors. The loss maximizes the mutual information between learned representations and device identity. For a batch of N signals, the model computes similarity scores and applies a softmax cross-entropy loss, effectively performing (N-1)-way discrimination per positive pair.

A critical application in RF fingerprinting is learning representations invariant to channel effects while remaining sensitive to hardware impairments. Domain-adversarial contrastive learning extends the framework by:

• Training a feature extractor to confuse a domain classifier that predicts channel conditions
• Ensuring the learned fingerprint depends on transmitter hardware, not propagation environment
• Enabling reliable device identification across varying multipath, distance, and noise conditions without retraining

Standard contrastive learning can struggle when devices have subtly different impairments. Hard negative mining identifies negative samples that are deceptively close to the anchor in embedding space—typically signals from transmitters of the same model and manufacturer. By emphasizing these difficult cases during training, the model learns to detect microscopic hardware variations such as minor I/Q imbalance differences or nearly identical phase noise profiles that distinguish otherwise identical devices.

Modern contrastive frameworks like SimCLR and MoCo employ a two-stage architecture:

• A base encoder (typically a ResNet or temporal convolutional network) processes raw I/Q samples into intermediate representations
• A projection head (a small MLP) maps these to a lower-dimensional space where contrastive loss is applied After training, the projection head is discarded, and the frozen base encoder serves as a general-purpose feature extractor for downstream tasks like device authentication, emitter classification, or anomaly detection.

A loss function that operates on triplets of samples: an anchor, a positive sample (same device), and a negative sample (different device). The objective is to ensure the distance between the anchor and positive is less than the distance between the anchor and negative by a specified margin.

• Enforces a relative distance constraint in the embedding space
• Critical hyperparameter: the margin value, which controls separation strictness
• Hard negative mining selects the most challenging different-device samples to accelerate convergence

A neural architecture where two or more identical subnetworks share weights and process different input samples in parallel. Each subnetwork produces an embedding, and a distance metric is computed between them.

• Weight sharing ensures identical feature extraction for both inputs
• Commonly paired with contrastive loss or triplet loss for training
• During inference, only one branch is needed to generate the fingerprint embedding for a new signal

Information Noise-Contrastive Estimation treats representation learning as a classification problem. Given a positive pair and a set of negative samples, the model must identify the true positive among all candidates.

• Based on mutual information maximization between related views
• Uses a softmax over similarity scores, scaled by a temperature parameter
• Forms the theoretical basis for SimCLR and other modern contrastive frameworks adapted for RF data

Creating valid positive pairs from a single transmitter's signal is a core design challenge. Techniques include applying signal augmentations that preserve device identity while varying channel conditions.

• Additive white Gaussian noise injection at varying SNR levels
• Simulated multipath fading and Doppler shift perturbations
• Time-shifting, cropping, and phase rotation of the raw IQ samples
• The augmentation policy must not destroy the subtle hardware impairments that constitute the fingerprint

A training strategy that prioritizes difficult negative examples — signals from different devices that the current model embedding incorrectly places close to the anchor. This prevents the model from learning trivial separations.

• Selects negatives with the highest similarity to the anchor in the current batch
• Essential for distinguishing transmitters with similar hardware impairments (same model, same manufacturing lot)
• Can be implemented online within a batch or offline using a memory bank of embeddings

The ultimate goal of contrastive learning for RF fingerprinting: an embedding space where distance corresponds to transmitter identity, not channel conditions. Domain-adversarial training is often combined with contrastive objectives to achieve this.

• Intra-device variance (caused by channel effects) is minimized
• Inter-device variance (caused by hardware impairments) is maximized
• Enables reliable authentication even when a device moves between environments