Feature Embedding

Embeddings from the same physical transmitter must cluster tightly together in the latent space, regardless of minor variations in the transmitted data payload. This requires the neural network to be invariant to modulation content while remaining sensitive to hardware-specific impairments. A low intra-class variance ensures that a single device consistently maps to a small, well-defined region, enabling reliable authentication with minimal false rejections.

Embeddings from different transmitters of the same make and model must be widely separated in the vector space. This is achieved through metric learning objectives like Triplet Loss or Contrastive Learning, which explicitly penalize overlapping clusters. High inter-class distance is critical for distinguishing between physically identical devices, where the only differentiating factors are microscopic manufacturing variances in analog components.

A robust embedding must remain stable despite multipath fading, Doppler shift, and noise. The latent representation should encode the transmitter's intrinsic hardware signature while filtering out environmental distortion. Techniques like Domain Adaptation and channel-robust data augmentation train the model to disentangle device-specific features from channel-specific artifacts, ensuring consistent operation in dynamic environments.

The latent space should be smooth and continuous, where interpolating between two embeddings produces a semantically meaningful intermediate representation. This property, enforced by Variational Autoencoders (VAEs) , ensures that small perturbations in the input signal do not cause discontinuous jumps in the embedding. A smooth manifold allows for reliable anomaly detection, as counterfeit or spoofed signals will map to low-density regions far from legitimate clusters.

The embedding dimension should be the minimum necessary to capture the intrinsic degrees of freedom of the transmitter hardware. Overly large embeddings waste storage and compute, while overly compressed embeddings lose discriminative information. The bottleneck layer of an autoencoder or the final projection layer of a Siamese network is carefully tuned to balance compactness with expressiveness, often targeting 64 to 256 dimensions for RF applications.

The embedding space must structurally support unknown emitter rejection. Rather than collapsing all inputs into known classes, the model should map unknown or rogue transmitters to low-density regions of the latent space. This is achieved by calibrating the embedding with Extreme Value Theory (EVT) to model the boundary of known device clusters, enabling the system to flag novel emitters that fall outside the expected distribution.

The compressed, lower-dimensional bottleneck representation where raw IQ samples or spectrograms are mapped. In a well-trained model, this space disentangles the factors of variation: device identity becomes a dominant organizing axis while channel artifacts are suppressed.

• Dimensionality typically ranges from 64 to 512 for RF applications
• Euclidean distance or cosine similarity in this space serves as the device similarity metric
• A smooth, continuous latent space enables interpolation between known signatures