Triplet Loss

The loss function operates on three distinct inputs simultaneously:

• Anchor (a): The reference sample, such as an RF fingerprint from a specific device.
• Positive (p): A sample belonging to the same class as the anchor (e.g., another transmission from the same device).
• Negative (n): A sample from a different class (e.g., a transmission from an impersonator or different device). The network processes all three through shared weights (a Siamese architecture) to generate embeddings.

A critical hyperparameter that defines the minimum distance enforced between positive and negative pairs relative to the anchor.

• The objective is: ||f(a) - f(p)||² + α < ||f(a) - f(n)||²
• Effect of α: A larger margin forces the model to create wider separation between classes, leading to more discriminative but potentially harder-to-train embeddings.
• Tuning: In RF fingerprinting, the margin must be carefully tuned to separate distinct devices without overfitting to channel-specific noise.

The strategy for selecting triplets during training is crucial for convergence and embedding quality.

• Random Triplets: Easy to generate but often yield trivial losses (the negative is already far away), leading to slow learning.
• Hard Negatives: Negatives that are closer to the anchor than the positive (||f(a) - f(n)|| < ||f(a) - f(p)||). These produce high loss and drive meaningful weight updates.
• Semi-Hard Negatives: Negatives that are farther than the positive but still within the margin. This is the standard strategy for stable training, avoiding collapsed representations.

In wireless fingerprinting, Triplet Loss is specifically engineered to learn channel-invariant features:

• Positive pairs are often formed by applying different synthetic channel distortions (multipath fading, noise) to the same device's clean signal. This forces the model to ignore channel effects.
• Negative pairs are formed from different devices under similar channel conditions, forcing the model to focus on hardware-specific impairments like I/Q imbalance or power amplifier non-linearity.
• This directly supports domain generalization by structuring the latent space around device identity, not propagation environment.

Several formulations exist to improve upon the standard hinge loss:

• Batch Hard Loss: For each anchor in a batch, selects the hardest positive and hardest negative. Maximizes discriminative power but is sensitive to label noise.
• Lifted Structured Loss: Considers all pairwise distances within a mini-batch, not just triplets, to exploit the full structural information.
• N-Pair Loss: Generalizes triplet loss by pushing an anchor against multiple negatives from distinct classes simultaneously, accelerating convergence.
• Angular Loss: Adds a geometric constraint on the angle at the negative point, providing scale invariance and more stable training.

Triplet Loss is foundational for few-shot device enrollment, where only a handful of transmissions are available to register a new IoT device.

• The model is pre-trained on a large set of known devices to learn a general embedding space.
• During enrollment, the few available samples serve as anchors. Authentication is performed by measuring the Euclidean distance between the enrolled anchor embeddings and a new live sample.
• If the distance falls below a calibrated threshold (derived from the margin α), the device is authenticated. This avoids retraining the entire network for each new device.