Higher-Order Statistics Classification

HOS classification relies on third-order (skewness) and fourth-order (kurtosis) cumulants to capture the shape of a signal's probability distribution. Unlike second-order statistics (variance, autocorrelation), cumulants are blind to Gaussian noise, meaning they can isolate non-Gaussian signal components even at low signal-to-noise ratios.

• Third-order cumulants detect asymmetry in the signal distribution, distinguishing between symmetric modulations like QPSK and asymmetric ones like 8-PAM.
• Fourth-order cumulants measure the peakedness of the distribution, separating sub-Gaussian signals (e.g., PSK) from super-Gaussian signals (e.g., multi-carrier OFDM).
• Cumulants of order greater than two are theoretically zero for Gaussian processes, providing inherent noise rejection.

HOS features form a discriminative fingerprint for digital modulation schemes. Each modulation family (PSK, QAM, FSK) exhibits a unique higher-order cumulant signature that remains consistent across varying symbol rates and carrier frequencies.

• M-PSK signals: Fourth-order cumulant values are negative and scale predictably with modulation order M.
• M-QAM signals: Cumulant values transition from negative to positive as constellation density increases, enabling separation of 16-QAM from 64-QAM.
• FSK signals: Exhibit distinct cumulant patterns due to their constant envelope property, differentiating them from linear modulations.
• This enables blind modulation classification without prior demodulation or timing recovery.

In real-world electromagnetic environments, interference often exhibits impulsive, heavy-tailed distributions that violate Gaussian assumptions. HOS classification thrives in these conditions because it explicitly models non-Gaussianity rather than treating it as noise.

• Atmospheric noise from lightning and man-made impulsive noise from switching electronics produce distributions with heavy tails that HOS features can characterize.
• Co-channel interference from other digital transmitters creates mixture distributions whose cumulant signatures reveal the number and types of interfering sources.
• HOS methods maintain classification integrity where second-order techniques fail catastrophically due to noise model mismatch.

The frequency-domain counterparts of third- and fourth-order cumulants are the bispectrum and trispectrum, which reveal phase coupling and nonlinearities invisible to the power spectrum.

• The bispectrum (B(f_1, f_2)) detects quadratic phase coupling between frequency components, a hallmark of nonlinear signal generation and certain jamming waveforms.
• The trispectrum captures cubic nonlinearities and provides a richer feature space for classifying complex interference patterns.
• These higher-order spectra are zero for Gaussian processes, making them ideal for isolating structured signals from colored Gaussian noise backgrounds.
• Applications include identifying nonlinear amplifier distortion and distinguishing between naturally occurring and intentionally generated interference.

Accurate HOS estimation requires larger sample sizes than second-order statistics due to higher variance in higher-order moment estimators. Practical systems must balance statistical reliability against latency constraints.

• The variance of k-th order cumulant estimates scales as (O(N^{-1})) for large sample sizes N, but with larger constant factors for higher orders.
• Adaptive estimation techniques use recursive updates to refine cumulant estimates as new samples arrive, enabling real-time classification.
• Robust estimators like trimmed means and median-based cumulants mitigate the impact of outlier samples in heavy-tailed noise.
• Typical implementations require 1,000–10,000 samples for reliable fourth-order cumulant estimation, translating to sub-millisecond processing at modern sampling rates.

Modern HOS classification pipelines combine handcrafted cumulant features with neural network backends to leverage both domain knowledge and learned representations.

• Cumulant vectors serve as compact, interpretable input features to fully connected or recurrent neural networks, reducing the dimensionality compared to raw IQ samples.
• Hybrid architectures use HOS features as one branch of a multi-input network, fused with spectrogram features from a CNN branch for complementary signal characterization.
• Autoencoder pre-training on HOS features learns compressed representations that are robust to channel impairments before fine-tuning for classification.
• This approach combines the noise rejection of HOS with the discriminative power of deep learning, achieving state-of-the-art performance in contested spectrum environments.

Exploits the periodic statistical properties of modulated signals for robust classification. Unlike HOS, which focuses on higher-order moments, cyclostationary analysis examines the spectral correlation function to identify unique cycle frequencies inherent to each modulation scheme. This technique excels in low signal-to-noise ratio (SNR) environments where traditional energy detection fails, making it complementary to cumulant-based methods for distinguishing between interference sources with overlapping power spectra.

A classification paradigm where models identify known signal types while flagging previously unseen interference patterns. Traditional closed-set classifiers force every input into a known category, creating dangerous misclassifications when novel jammers appear. Open-set techniques use extreme value theory on HOS feature spaces to model the probability that a sample belongs to an unknown class, enabling safe rejection of unfamiliar waveforms in contested electromagnetic environments.

The hardening of RF machine learning models against evasion attacks where intelligent jammers subtly manipulate waveforms to fool classifiers. Attack vectors include:

• Fast Gradient Sign Method (FGSM): Perturbing IQ samples to cross decision boundaries
• Carlini-Wagner attacks: Minimal distortion perturbations targeting specific misclassifications Defense strategies involve adversarial training with HOS-constrained perturbations and certified robustness bounds using cumulant-based Lipschitz constant estimation, ensuring classifiers remain reliable against adaptive adversaries.

A mathematical representation capturing spatial correlation between signals received at multiple antennas. For an M-element array, the covariance matrix R = E[xx^H] encodes both second-order statistics (diagonal elements) and cross-correlations (off-diagonal elements). When extended to higher orders, the cumulant tensor generalizes this concept, enabling classification of non-Gaussian interference sources through their multidimensional statistical signatures, particularly useful for separating co-channel interferers with identical power spectra.

The application of feature attribution methods to decode complex RF classification decisions for human analysts. Techniques include:

• SHAP (SHapley Additive exPlanations): Quantifies each cumulant's contribution to a classification
• Integrated Gradients: Highlights which time-frequency regions most influenced the decision
• LIME: Creates interpretable local surrogate models around specific predictions XAI bridges the gap between high-accuracy HOS-based classifiers and operator trust, enabling spectrum managers to understand why a signal was flagged as hostile interference.