Dimensionality Reduction

The foundational linear technique that transforms correlated feature variables into a set of linearly uncorrelated principal components. PCA identifies the directions (eigenvectors) of maximum variance in the data, allowing engineers to discard low-variance components that represent noise rather than signal.

• Mechanism: Eigendecomposition of the covariance matrix
• Output: Ranked components by explained variance ratio
• Use Case: Reducing a 1024-point bispectrum to its top 50 principal components for real-time SEI classification
• Limitation: Assumes linear relationships; may fail to capture non-linear hardware impairment interactions

A non-linear technique designed primarily for visualization of high-dimensional data in 2D or 3D spaces. t-SNE preserves local structure by modeling pairwise similarities as probabilities, making it excellent for exploratory analysis of whether different emitters form distinct clusters in feature space.

• Mechanism: Minimizes KL divergence between high-dimensional and low-dimensional probability distributions
• Key Parameter: Perplexity, which balances attention between local and global structure
• Use Case: Visualizing the separability of 20 WiFi device fingerprints before training a classifier
• Limitation: Non-deterministic output; computationally expensive for large datasets; not suitable as a preprocessing step for downstream models

A modern non-linear technique that competes with t-SNE for visualization quality while better preserving the global structure of the data. UMAP is built on rigorous mathematical foundations from Riemannian geometry and topological data analysis, making it faster and more scalable than t-SNE.

• Mechanism: Constructs a fuzzy topological representation of the high-dimensional data, then optimizes a low-dimensional embedding to be as structurally similar as possible
• Advantage: Significantly faster than t-SNE; can scale to millions of data points
• Use Case: Projecting a large dataset of RF-DNA feature vectors into 3D to identify unknown emitter clusters in an open set recognition task
• Key Parameter: n_neighbors, controlling the balance between local and global structure preservation

A deep learning approach where a neural network is trained to reconstruct its own input through a bottleneck layer. The bottleneck (latent space) learns a compressed, non-linear representation of the input features. Unlike PCA, autoencoders can capture complex, non-linear relationships in hardware impairment data.

• Architecture: Encoder compresses input → Bottleneck (latent vector) → Decoder reconstructs
• Variants:
  • Sparse Autoencoders: Add sparsity constraints to learn interpretable features
  • Denoising Autoencoders: Train on corrupted inputs to learn robust representations
  • Variational Autoencoders (VAEs): Learn a probabilistic latent space for generative modeling
• Use Case: Compressing a 2048-point cyclostationary feature vector into a 64-dimensional latent embedding that serves as the device's fingerprint for a Siamese authentication network

A supervised dimensionality reduction technique that projects data onto a lower-dimensional space while maximizing class separability. Unlike PCA, which is unsupervised and focuses solely on variance, LDA explicitly uses device labels to find the projection that best discriminates between different emitters.

• Mechanism: Maximizes the ratio of between-class variance to within-class variance (Fisher's criterion)
• Output: At most (C-1) dimensions, where C is the number of emitter classes
• Use Case: Reducing a high-dimensional feature vector to a 2D space where 15 different transmitter models are maximally separated for a k-Nearest Neighbor classifier
• Limitation: Assumes normally distributed classes with equal covariance; may underperform when these assumptions are violated by complex hardware impairment distributions

A statistical approach that selects the most informative subset of original features rather than transforming them. Mutual information measures the dependency between each feature and the emitter identity, allowing engineers to discard irrelevant or redundant features while maintaining physical interpretability.

• Mechanism: Ranks features by their mutual information score with the target class label
• Advantage: Preserves the physical meaning of features (e.g., "carrier frequency offset" remains interpretable, unlike PCA components)
• Use Case: Selecting the top 20 most discriminative features from a 200-dimensional set of I/Q imbalance, phase noise, and EVM measurements for a lightweight edge deployment
• Limitation: Does not account for feature interactions; a feature individually weak may be strong in combination with others

A linear transformation that identifies the orthogonal directions (principal components) of maximum variance in the feature space. In RF fingerprinting, PCA projects a high-dimensional feature vector—such as bispectrum coefficients or wavelet energy maps—onto a lower-dimensional subspace, discarding dimensions dominated by noise while retaining the discriminative hardware impairment signatures that separate emitters.

• Unsupervised; requires no class labels
• First principal component captures the largest variance direction
• Often used as a pre-processing step before classification with SVMs or k-NN

A non-linear dimensionality reduction technique optimized for visualizing high-dimensional data in 2D or 3D scatter plots. t-SNE preserves local neighborhood structure, making it invaluable for exploratory analysis of RF fingerprint clusters—revealing whether different devices form separable groupings in the feature space before committing to a classifier architecture.

• Excels at revealing natural clusters and outliers
• Perplexity parameter controls the balance between local and global structure
• Computationally intensive; typically applied to pre-reduced feature sets

A supervised dimensionality reduction method that projects features onto axes that maximize class separability rather than total variance. For emitter identification, LDA finds the linear combinations of features that best separate known transmitters while minimizing within-class scatter, making it directly aligned with the authentication objective.

• Requires labeled training data for each known device
• Produces at most (C-1) dimensions for C classes
• Often outperforms PCA when class labels are available during enrollment

A neural network architecture trained to reconstruct its input through a bottleneck layer of reduced dimensionality. The bottleneck activations form a compressed, non-linear representation of the RF fingerprint. Variational autoencoders (VAEs) extend this by learning a probabilistic latent space, enabling generative modeling of device signatures and robust feature extraction under noisy channel conditions.

• Learns non-linear manifolds that PCA cannot capture
• Bottleneck size controls compression ratio
• Decoder enables reconstruction for signal denoising

Two distinct paradigms for reducing dimensionality. Feature selection retains a subset of original features (e.g., selecting specific spectral moments or EVM components) based on statistical relevance scores. Feature extraction—including PCA and autoencoders—creates entirely new synthetic features. In RF fingerprinting, selection offers interpretability (knowing which hardware impairment matters), while extraction often yields superior classification accuracy.

• Selection: mutual information, recursive feature elimination
• Extraction: PCA, LDA, autoencoders, UMAP
• Hybrid approaches apply selection after extraction

The phenomenon where classifier performance degrades as feature dimensionality increases beyond an optimal point, driven by data sparsity in high-dimensional space. A bispectrum or wavelet transform can produce thousands of coefficients, but with limited training captures per device, the ratio of samples to features collapses. Dimensionality reduction directly mitigates this curse, improving generalization and reducing overfitting.

• Distance metrics lose discriminative power in high dimensions
• Required training samples grow exponentially with dimensions
• Reduction to 10–50 dimensions is typical for SEI systems