Contrastive Learning

The engine of contrastive learning is the creation of positive pairs—two views of the same underlying data point. In RF fingerprinting, a single transmitted waveform is augmented with synthetic channel distortions (multipath fading, Doppler shift, noise) to create multiple views. The model learns that these views share a common device identity, forcing it to ignore channel-specific artifacts and focus on hardware-intrinsic impairments.

• Augmentation strategies: Additive white Gaussian noise, simulated multipath profiles, frequency offsets
• Key principle: Maximize mutual information between augmented views of the same emitter
• Outcome: Representations invariant to channel conditions but discriminative of device identity

Equally critical is the negative pair—two views from different emitters. The loss function explicitly penalizes the model if it maps different devices close together. This repulsive force prevents representation collapse, where the encoder maps all inputs to a trivial constant vector. In large-batch training, negative samples are drawn from other emitters in the batch, creating a dynamic discrimination task.

• Hard negative mining: Selecting negatives that are currently close in embedding space to accelerate learning
• Memory banks: Storing representations from previous batches to increase the pool of negatives without ballooning batch size
• Uniformity property: The embedding space is encouraged to distribute features uniformly on the hypersphere

The InfoNCE (Noise Contrastive Estimation) loss is the dominant objective function. It frames the task as a categorical classification: given one positive and N negatives, identify the correct pair. A temperature parameter (τ) controls the concentration of the similarity distribution. Lower temperatures sharpen the distribution, making the model focus on hard negatives; higher temperatures soften it, providing a more uniform gradient signal.

• Formula structure: Softmax over cosine similarities, scaled by 1/τ
• Temperature tuning: Critical hyperparameter balancing uniformity and tolerance to semantically similar negatives
• Relation to mutual information: InfoNCE provides a lower bound on the mutual information between positive views

Contrastive architectures employ a projection head—a small multi-layer perceptron appended to the encoder backbone. Contrastive loss is applied to the output of this head, not the encoder directly. After training, the head is discarded, and the encoder's penultimate layer is used for downstream fingerprinting. This design prevents the loss from discarding information useful for device discrimination but irrelevant for the contrastive pretext task.

• Architecture: Encoder → Projection MLP → Contrastive loss
• Discard after training: The projection head acts as a disposable information filter
• Benefit: Preserves richer device-specific features in the encoder's representation

Many state-of-the-art contrastive frameworks employ a momentum encoder—a slowly evolving copy of the main encoder updated via exponential moving average (EMA). This provides stable, consistent target representations for the positive pair, preventing the training signal from oscillating as the main encoder updates rapidly. In RF contexts, this stability is crucial when learning subtle hardware impairments amid noisy channel variations.

• Update rule: θ_momentum = m·θ_momentum + (1-m)·θ_encoder, with m typically 0.999
• Purpose: Decouples the target representation from the rapidly updating online encoder
• Asymmetric architecture: Online encoder processes one view; momentum encoder processes the other

When applied to RF fingerprinting, contrastive learning directly addresses the domain shift problem. By constructing positive pairs that share device identity but differ in channel conditions, the model learns to disentangle hardware impairments (stable, device-specific) from channel effects (variable, environment-specific). This enables a single model to authenticate devices across diverse propagation environments without per-environment calibration.

• Positive pair: Same device, different synthetic channel profiles
• Negative pair: Different devices, any channel conditions
• Result: Embedding clusters organized by device identity, not channel characteristics

A metric learning loss function that operates on triplets of data points: an anchor, a positive sample (same device, different channel), and a negative sample (different device). The loss minimizes the distance between the anchor and positive while enforcing a margin of separation from the negative. In RF fingerprinting, this directly trains the embedding space to cluster by device identity rather than channel conditions.

A neural architecture consisting of two or more identical subnetworks that share weights and process distinct input samples. During training, the network learns a similarity function between pairs of RF bursts. For contrastive fingerprinting, one branch processes a clean reference signal while the other processes a channel-impaired version, forcing the shared encoder to extract channel-invariant features.

A technique that combines contrastive objectives with an adversarial domain classifier. While the main network learns to separate device identities, a gradient reversal layer forces the feature extractor to maximize domain classifier error. This ensures the learned representations are discriminative for device ID but indistinguishable across different RF channel environments, directly implementing channel-robustness.

A regularization strategy that artificially expands the training dataset by applying label-preserving transformations. For contrastive RF learning, this includes adding synthetic multipath profiles, Doppler shifts, and noise. By treating augmented versions of the same signal as positive pairs, the model learns to ignore channel-induced variations and focus on the underlying hardware impairments.

The process of separating a learned representation into independent, interpretable factors. In channel-robust fingerprinting, the goal is to isolate device-specific features (DAC non-linearity, I/Q imbalance) from channel-specific features (multipath delay, fading). Contrastive learning with carefully designed positive/negative pairs encourages this separation without explicit labels.

A kernel-based statistical measure of the distance between two probability distributions. Used as a regularization term alongside contrastive loss, MMD aligns the feature distributions of signals from different RF channels. By minimizing MMD between domain representations, the model ensures that the embedding space structure depends on device identity, not propagation conditions.