Contrastive Learning

The core mechanism relies on creating positive pairs—two augmented views of the same RF signal that the network must learn to identify as similar. Augmentations include:

• Additive Gaussian noise injection to simulate varying SNR conditions
• Frequency offset application to mimic oscillator drift
• Time shifting to account for burst timing jitter
• Phase rotation to handle carrier phase ambiguity

These transformations preserve the underlying hardware fingerprint while teaching the encoder invariance to channel and receiver artifacts.

The InfoNCE (Noise Contrastive Estimation) loss function is the mathematical engine driving representation learning. For a batch of N signals, it treats the positive pair as the only correct match among N-1 negative distractors.

• The temperature parameter τ controls the concentration of the distribution—lower values create sharper distinctions between similar and dissimilar emitters
• The loss maximizes mutual information between augmented views while minimizing it between different devices
• This formulation directly optimizes for inter-class separability and intra-class compactness in the embedding space

The SimCLR architecture, originally designed for computer vision, has been adapted for complex-valued RF signals:

• A projection head (typically a 2-3 layer MLP) maps encoder outputs to a space where contrastive loss is applied, then is discarded after training
• The retained base encoder—often a 1D ResNet or temporal convolutional network—processes raw IQ samples directly
• Larger batch sizes (512-4096) are critical, providing more negative examples and improving the quality of learned representations
• Layer normalization replaces batch normalization to handle the wide dynamic range of RF power levels

An alternative to contrastive loss, Barlow Twins operates on the cross-correlation matrix between embeddings of augmented signal pairs:

• The objective drives the matrix toward the identity matrix, making each embedding component invariant to augmentations while decorrelating different components
• This prevents dimensional collapse, where the encoder maps all signals to a trivial constant vector
• Particularly effective for RF data where the number of unique emitters is unknown, as it doesn't require explicit negative pair selection
• Naturally resistant to class collision issues that plague contrastive methods when two different emitters are accidentally paired as negatives

MoCo (Momentum Contrast) maintains a dynamic dictionary of encoded signal representations using a momentum-updated encoder:

• A queue stores thousands of negative sample embeddings, decoupling batch size from the number of negatives
• The momentum encoder is a slowly-evolving copy of the main encoder, ensuring consistent key representations over time
• Critical for RF fingerprinting scenarios with hundreds of known emitters, where a single batch cannot contain sufficient negative diversity
• Enables continual learning as new emitter signatures can be added to the queue without full retraining

After pre-training, the frozen or fine-tuned encoder serves as a powerful feature extractor for supervised tasks:

• Linear evaluation: A simple classifier trained on frozen embeddings often achieves near-supervised performance with only 10-20% labeled data
• Few-shot identification: With as few as 5-10 examples per emitter, prototypical networks built on contrastive embeddings can achieve >95% accuracy
• The learned representations naturally exhibit channel robustness, having been trained to ignore the augmentation-based channel variations
• Embeddings can be directly used for open-set recognition by thresholding distances in the learned metric space

A dual-branch architecture that processes pairs of RF signals through identical weight-shared subnetworks to learn a similarity metric. During training, one branch receives an anchor sample while the other receives either a positive (same emitter) or negative (different emitter) sample. The network outputs two embedding vectors, and the distance between them is minimized for positive pairs and maximized for negative pairs. This architecture is foundational for one-shot learning in emitter identification, enabling classification of devices seen only once during enrollment.

A metric learning objective that organizes the embedding space using triplets of samples: an anchor, a positive (same emitter), and a negative (different emitter). The loss function enforces that the distance between the anchor and positive is smaller than the distance between the anchor and negative by at least a specified margin. This creates tightly clustered, well-separated emitter representations. Key considerations include:

• Hard negative mining selects the most confusable negative samples to accelerate convergence
• Semi-hard triplets avoid training collapse by ignoring trivial negatives
• The margin hyperparameter controls intra-class compactness versus inter-class separation

A few-shot learning architecture that computes a prototype representation for each known emitter by averaging the embeddings of its support set examples. Classification of a new query signal is performed by finding the nearest prototype in the embedding space using Euclidean distance. This approach is particularly effective for RF fingerprinting because:

• It naturally handles open-set recognition by thresholding distance to the nearest prototype
• New emitters can be enrolled without retraining by simply computing their prototype
• The averaging operation provides inherent noise resilience for signals captured under varying SNR conditions

A neural network component used in domain-adversarial training to achieve channel-invariant fingerprint representations. During backpropagation, the gradient reversal layer multiplies the gradient by a negative scalar, forcing the feature extractor to learn representations that are discriminative for emitter identity but uninformative for channel conditions. This technique directly addresses the channel robustness problem in RF fingerprinting, where a model trained in one environment fails when deployed in another. The architecture consists of:

• A feature extractor shared across domains
• A label classifier for emitter identity
• A domain classifier with the gradient reversal layer that attempts to predict the capture environment

A transfer learning paradigm that aligns the feature distributions of RF fingerprints captured under different channel conditions or receiver hardware. Without adaptation, a model trained on line-of-sight captures will catastrophically fail on multipath-rich indoor signals. Key techniques include:

• Maximum Mean Discrepancy (MMD) minimization to match source and target distributions
• Adversarial domain confusion using a discriminator network
• Correlation alignment to match second-order statistics between domains This enables practical deployment where training data is collected in controlled environments but inference occurs in the field.

A classification paradigm where the model must simultaneously identify known emitters and reject unknown devices whose fingerprints were absent from training. Contrastive learning naturally supports open-set recognition because the embedding space is structured by similarity rather than fixed class boundaries. Implementation approaches include:

• Extreme Value Theory to model the distribution of known-class distances and set rejection thresholds
• Angular margin penalties to create compact known-class clusters with large inter-class gaps
• Distance-based detection where queries exceeding a calibrated threshold from all prototypes are flagged as rogue devices