Uniform Manifold Approximation and Projection (UMAP)

UMAP operates in two phases:

• Graph Construction: It builds a weighted k-neighbor graph by assuming the data lies on a locally connected Riemannian manifold. The local metric is scaled by a local connectivity constraint, ensuring uniform coverage.
• Low-Dimensional Optimization: It then optimizes a low-dimensional layout using cross-entropy as a loss function between the high-dimensional and low-dimensional fuzzy topological representations. This rigorous mathematical basis provides a more scalable and theoretically grounded alternative to purely probabilistic methods.

A non-linear dimensionality reduction technique that excels at preserving local data structure. t-SNE converts high-dimensional Euclidean distances into conditional probabilities representing similarities, then minimizes the Kullback-Leibler divergence between the high- and low-dimensional distributions.

• Key difference from UMAP: t-SNE prioritizes local neighborhood preservation over global structure
• Computational cost: O(n²) complexity makes it significantly slower than UMAP on large datasets
• Common use case: Visualizing feature embeddings of emitter clusters in 2D or 3D scatter plots
• Limitation: Stochastic nature means different runs produce different layouts; hyperparameter tuning (perplexity) is critical

The process of mapping high-dimensional IQ data or spectrograms into a lower-dimensional vector space where semantically similar device signatures cluster together. Embeddings serve as the input to UMAP for visualization and as the foundation for similarity-based authentication.

• Dimensionality: Typical RF embeddings range from 64 to 512 dimensions before UMAP projection
• Training objective: Triplet loss or contrastive learning ensures same-device signals map to nearby vectors
• Downstream use: Embedding vectors feed into cosine similarity calculations for few-shot device enrollment
• Quality metric: Silhouette score measures how well-separated device clusters are in the embedding space

The compressed, abstract representation of input data learned by a neural network's bottleneck layer. In Specific Emitter Identification, the latent space captures the intrinsic factors of variation related to hardware impairments while discarding nuisance variables like modulation content.

• Disentanglement: Well-trained latent spaces separate device identity from channel effects and noise
• Visualization: UMAP projects the latent space into 2D to reveal natural clustering of transmitter populations
• Anomaly detection: Unknown or spoofed devices appear as outliers far from known clusters in the latent representation
• Generative modeling: Variational Autoencoders impose a probabilistic structure on the latent space, enabling synthetic signal generation

A classification methodology that extends beyond closed-world assumptions by not only identifying known emitter classes but also reliably detecting and rejecting unknown transmitters not seen during training. UMAP visualizations are critical for validating open set performance.

• Core challenge: Standard SoftMax classifiers force all inputs into known classes, even rogue devices
• Solution: OpenMax replaces SoftMax with a calibrated rejection mechanism using Extreme Value Theory
• UMAP validation: Visual inspection of latent space reveals whether unknown devices form distinct clusters or overlap with known classes
• Operational metric: Area Under the ROC Curve for open set detection measures the trade-off between known-class accuracy and novelty rejection

A self-supervised learning paradigm that trains models to pull representations of similar signal samples together and push dissimilar ones apart in the embedding space. This directly optimizes the cluster structure that UMAP later visualizes.

• Training mechanism: Positive pairs (same device, different captures) are attracted; negative pairs (different devices) are repelled
• Loss functions: Triplet loss and InfoNCE loss are the dominant objectives for RF fingerprinting
• UMAP synergy: Contrastive embeddings produce cleaner, more interpretable UMAP projections with well-separated device clusters
• Few-shot advantage: Contrastively pre-trained models require minimal enrollment samples for new device registration

A subfield of transfer learning focused on mitigating the distribution shift between training data and operational data caused by varying channel environments. Without adaptation, UMAP visualizations reveal catastrophic cluster drift when models deploy to new RF conditions.

• Problem: A fingerprinting model trained in a lab anechoic chamber fails in a multipath-rich urban environment
• Techniques: Adversarial domain adaptation and maximum mean discrepancy minimization align source and target distributions
• UMAP diagnostic: Plotting source and target embeddings together reveals domain gaps as systematic offsets between clusters
• Channel-robust learning: Contrastive objectives that treat channel variation as a nuisance variable produce domain-invariant features