Contrastive Learning

The engine of contrastive learning is the creation of positive pairs—two augmented views of the same underlying sample. For RF fingerprinting, a single IQ sample from a transmitter is transformed using domain-specific augmentations:

• Additive Gaussian noise injection
• Frequency shift simulation
• Time cropping or translation
• Amplitude scaling The network is trained to recognize these distorted versions as semantically identical, forcing it to learn features invariant to channel impairments while remaining sensitive to the hardware-specific fingerprint.

Equally critical to pulling positives together is pushing negative pairs apart. A negative pair consists of signals from two different physical transmitters. The loss function explicitly penalizes the network when embeddings from distinct devices occupy nearby positions in the latent space. This creates a discriminative margin between device clusters. In practice, hard negative mining—selecting the most confusable impostor devices for a given emitter—significantly accelerates convergence and sharpens decision boundaries.

The de facto standard objective for contrastive learning is InfoNCE (Noise Contrastive Estimation). It frames the task as a categorical classification over a batch: given one positive example and K negative examples, the model must correctly identify the positive. The loss is:

• Low when the positive pair similarity score dominates all negative pair scores
• High when any negative embedding is closer than the positive This formulation directly maximizes the mutual information between different views of the same signal, learning representations that capture the underlying device-specific factors of variation.

An alternative to InfoNCE is triplet loss, which operates on explicit triplets of samples:

• Anchor: A reference signal from a specific device
• Positive: Another signal from the same device
• Negative: A signal from a different device The loss enforces the constraint: distance(anchor, positive) + margin < distance(anchor, negative). This margin hyperparameter controls how tightly clusters form. Triplet-based training is particularly effective for open-set device identification, where clear inter-class separation is essential for rejecting unknown emitters.

A critical architectural insight from SimCLR is the use of a non-linear projection head during training. The encoder network feeds into a small multi-layer perceptron (MLP) that projects representations to a lower-dimensional space where the contrastive loss is applied. After training, this head is discarded, and the encoder's output is used directly for downstream fingerprinting tasks. This prevents the loss from compressing out information useful for device discrimination, preserving hardware-specific features in the encoder's representation.

The temperature parameter (τ) in the contrastive loss controls the concentration of the similarity distribution. A low temperature (< 1.0) sharpens the distribution, making the model focus intensely on hard negatives—impostor devices with very similar fingerprints. A high temperature smooths the distribution, treating all negatives more uniformly. For RF fingerprinting, tuning τ is essential: too low causes instability from noisy RF samples, while too high fails to separate devices with subtle hardware differences like similar power amplifier non-linearity profiles.

The foundational loss function for contrastive learning 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 defined margin. This directly enforces the clustering of same-device RF fingerprints in the embedding space.

A twin-branch architecture where two identical subnetworks with shared weights process two distinct input samples in parallel. Each branch generates an embedding vector, and the contrastive loss is applied to the distance between these outputs. This weight-sharing guarantees that identical signals produce identical embeddings, making it ideal for verifying if two RF bursts originate from the same emitter.

A training strategy that selectively focuses on the most difficult negative samples—those from different devices whose embeddings are deceptively close to the anchor. By prioritizing these confusing imposters, the model learns a more robust and sharply defined decision boundary. This is critical for distinguishing emitters with very similar hardware impairments.

Noise Contrastive Estimation (InfoNCE) reframes the problem as a categorical classification task. Given an anchor and a set of samples containing one positive match and many negative distractors, the model must correctly identify the positive. This loss function maximizes the mutual information between the learned representations, scaling efficiently to very large batch sizes for robust RF fingerprinting.

The systematic application of realistic channel transformations to generate positive pairs for contrastive learning. Augmentations define what constitutes invariance for a fingerprint. Common RF augmentations include:

• Additive White Gaussian Noise (AWGN)
• Simulated Rayleigh or Rician fading
• Small, random Carrier Frequency Offset (CFO) shifts
• Time-domain cropping and resampling

The process of projecting high-dimensional RF fingerprint vectors into 2D or 3D for qualitative analysis using techniques like t-SNE or UMAP. A well-trained contrastive model produces a visualization where signals from the same device form tight, isolated clusters, while clusters from different devices are clearly separated. This is a primary diagnostic for evaluating model convergence and discriminative power.