Contrastive Learning for RF

The core mechanism creates multiple distorted versions of the same baseband I/Q signal through domain-specific augmentations. These include:

• Additive white Gaussian noise injection at varying SNR levels
• Random time shifts and cropping of I/Q sequences
• Channel impulse response convolution to simulate multipath
• Frequency offset rotation to mimic oscillator drift

The model learns that these perturbed views represent the same transmitter identity, forcing it to ignore channel artifacts and focus on hardware-intrinsic features.

The Normalized Temperature-scaled Cross Entropy loss operates on batches of signal pairs. For each positive pair (two augmented views of the same signal), the loss maximizes cosine similarity in embedding space. Simultaneously, it minimizes similarity for all negative pairs (views from different transmitters).

The temperature parameter controls concentration: lower values sharpen the distribution, penalizing hard negatives more aggressively. This is critical for separating transmitters with subtly different hardware impairments like near-identical power amplifier non-linearities.

A small multi-layer perceptron attached to the encoder output maps representations to a lower-dimensional space where contrastive loss is applied. This projection head is discarded after pre-training—only the encoder backbone is retained for downstream fingerprinting tasks.

The head prevents the representation from collapsing to trivial solutions by acting as a bottleneck that filters out augmentation-specific noise while preserving transmitter-discriminative information. Typical configurations use 2-3 layers with ReLU activation and output dimensions of 128-256.

Random negative sampling often yields easy negatives from transmitters with grossly different fingerprints. Hard negative mining deliberately selects confusable transmitter pairs—devices of the same model and manufacturing batch—that the current model embedding places close together.

By prioritizing these challenging examples during training, the model learns to discriminate microscopic hardware variations such as subtle I/Q imbalance differences of 0.1-0.5 dB between otherwise identical software-defined radios. This dramatically improves open-set rogue device detection.

A slowly-evolving copy of the main encoder, updated via exponential moving average rather than backpropagation, maintains a consistent dictionary of negative representations. This decouples the dictionary size from the mini-batch size, enabling large-scale contrastive pre-training on millions of unlabeled signal captures.

The momentum coefficient (typically 0.999) ensures the dictionary representations evolve smoothly, preventing the model from exploiting inconsistency shortcuts where negatives from different training steps have incompatible embeddings.

After self-supervised pre-training, the frozen encoder is evaluated by training only a linear classifier on top of the learned representations using a small labeled dataset. This protocol isolates the quality of the learned embeddings from the classifier's capacity.

Strong linear probe performance—achieving 95%+ emitter identification accuracy with only 10-20 labeled examples per transmitter—demonstrates that the contrastive objective has successfully captured physically meaningful RF fingerprint features without any label supervision during pre-training.

A margin-based loss function that structures the embedding space by enforcing a relative distance constraint between signal representations. For each training sample, an anchor (reference signal), positive (same transmitter, different capture), and negative (different transmitter) are selected. The loss penalizes embeddings where the anchor-positive distance exceeds the anchor-negative distance by less than a specified margin. This produces tightly clustered transmitter identities with clear inter-class separation.

The core loss function used in SimCLR-style contrastive learning adapted for RF signals. It treats each augmented pair from the same signal as a positive example and all other signals in the batch as negatives. Key components include:

• L2 normalization of embeddings to project onto a hypersphere
• Temperature parameter that controls the concentration of the distribution
• Large batch sizes to provide sufficient negative examples for discriminative learning

A critical component of contrastive pre-training that generates semantically invariant views of the same transmitter signal. Domain-specific augmentations include:

• Additive white Gaussian noise injection at varying SNR levels
• Frequency offset simulation to mimic oscillator drift
• Time shifting and cropping to remove absolute temporal dependencies
• Channel impulse response convolution to simulate multipath The choice of augmentations defines which nuisance factors the learned representation becomes invariant to.

A complementary technique that learns channel-invariant transmitter fingerprints by jointly training a feature extractor with a gradient reversal layer. While the main encoder learns to identify transmitters, a domain classifier attempts to predict the channel conditions (indoor, outdoor, vehicular). By reversing gradients during backpropagation, the encoder is forced to suppress channel-specific features while preserving transmitter-specific discriminative information. Often combined with contrastive objectives for robust open-set recognition.

A simplified self-supervised framework that eliminates the need for negative pairs or large batch sizes, making it practical for RF datasets with limited transmitter diversity. The architecture uses a stop-gradient operation on one branch of a Siamese network to prevent representational collapse. For RF applications, this enables effective pre-training on unlabeled spectrum captures where only a handful of known transmitters exist, learning robust features that transfer well to downstream SEI tasks.