Feature Vector Extraction

The feature vector must capture inter-device variance that is significantly larger than intra-device variance. This means the mathematical distance between vectors from different transmitters must be maximized, while vectors from the same transmitter remain tightly clustered. Features derived from power amplifier non-linearity or I/Q imbalance often provide high discriminability because these impairments are unique to each analog front-end. Without this property, a classifier cannot establish clear decision boundaries.

A feature vector must be invariant to the multipath fading and channel state information (CSI) of the transmission environment. If a feature changes significantly between a line-of-sight and a non-line-of-sight scenario, it is encoding the channel, not the device. Techniques like cyclostationary analysis exploit signal statistics that are inherently robust to dispersive channels. Alternatively, domain adaptation algorithms can be applied to transform features from a source environment to a target environment, decoupling the channel from the hardware signature.

The feature vector must be compact to avoid the curse of dimensionality and enable real-time processing. Raw I/Q samples can produce millions of dimensions, which is computationally prohibitive. Effective extraction applies dimensionality reduction techniques like Principal Component Analysis (PCA) or learns a low-dimensional embedding space using a Siamese network. A vector of 50-200 highly informative features is often more effective for classification than a sparse vector of thousands, reducing both latency and overfitting risk.

The extracted features must remain consistent over short to medium time horizons to prevent false rejections. However, they must also be paired with drift compensation mechanisms to account for slow, natural variations caused by temperature changes and component aging. A feature like carrier frequency offset (CFO) is highly discriminative but can drift with oscillator warm-up. An effective feature vector design either selects features with inherent long-term stability or includes a dynamic device signature baseline update protocol.

The extraction algorithm must be feasible for the target deployment platform, whether that is a cloud server, an FPGA, or a low-power edge AI processor. Complex transforms like the bispectrum provide rich, noise-resistant features but are computationally expensive. In contrast, statistical moments or wavelet domain fingerprint coefficients can be computed with significantly less overhead. The design must balance the signal processing gain against the available multiply-accumulate operations per second.

The feature vector must be resilient to the random additive white Gaussian noise (AWGN) inherent in any receiver. Features should be extracted from the deterministic signal components, not the stochastic noise floor. Higher-order statistics (HOS) like kurtosis are theoretically immune to Gaussian noise, making them powerful features. Similarly, spectral regrowth patterns, which are caused by deterministic non-linear compression, remain identifiable even at low signal-to-noise ratios where the ideal constellation is obscured.

A set of techniques used to compress a high-dimensional feature vector into a lower-dimensional space while preserving its identifying variance. Principal Component Analysis (PCA) is the most common method, projecting features onto orthogonal axes of maximum variance. Other techniques include t-Distributed Stochastic Neighbor Embedding (t-SNE) for visualization and Linear Discriminant Analysis (LDA) for maximizing class separability. This step is critical for reducing computational complexity and mitigating the curse of dimensionality in downstream classifiers.

The analysis of a signal's third-order (skewness) and fourth-order (kurtosis) statistical moments and their frequency-domain representations. Key techniques include:

• Bispectrum Analysis: Transforms a signal to reveal quadratic phase coupling, providing a noise-resistant feature space
• Trispectrum Analysis: Captures fourth-order cumulants for characterizing non-Gaussian emitter behavior
• Cumulant-based features are inherently immune to Gaussian noise, making them highly robust for low-SNR environments

A feature extraction method that applies a wavelet transform to decompose a signal into joint time-frequency representations. Unlike Fourier-based methods, wavelets provide multi-resolution analysis, capturing both transient events (like turn-on spikes) and steady-state characteristics simultaneously. Common approaches include Discrete Wavelet Transform (DWT) for compression and Continuous Wavelet Transform (CWT) for detailed scalogram visualization. This technique excels at isolating non-stationary impairment signatures that stationary transforms miss.

A signal processing technique that exploits the periodic statistical properties of modulated signals. Man-made communication signals exhibit cyclostationarity due to carrier frequencies, symbol rates, and guard intervals. The Spectral Correlation Function (SCF) maps these cyclic frequencies, generating features that are:

• Robust to stationary noise and interference
• Modulation-specific, aiding in joint classification and identification
• Separable from overlapping signals in dense spectrum environments

Features derived from analyzing errors in the in-phase and quadrature components of a modulated signal. Key extracted metrics include:

• Error Vector Magnitude (EVM): Aggregates multiple impairments into a single quality score
• I/Q Imbalance: Gain and phase mismatch between I and Q branches
• Origin Offset: DC offset caused by local oscillator leakage
• Quadrature Skew: Deviation from the ideal 90-degree phase separation These scalar features provide a compact, interpretable representation of modulator hardware imperfections.

The process of mapping extracted feature vectors into a high-dimensional space where semantic similarity equals Euclidean proximity. Deep learning architectures like Siamese Networks and Triplet Networks learn this mapping explicitly, ensuring that signals from the same device cluster tightly while different devices are pushed apart. Once constructed, device identity is verified by measuring cosine similarity or L2 distance between a probe vector and enrolled baselines. This approach enables efficient one-shot and few-shot authentication.