Cumulant-Based Deep Feature Extraction

Higher-order cumulants (order ≥ 3) are theoretically zero for Gaussian processes, making them inherently immune to additive white Gaussian noise. This property allows the feature vector to capture only the non-Gaussian modulation structure, dramatically improving classification performance at low SNR where raw IQ-based deep learning models struggle.

• Third-order and fourth-order cumulants suppress Gaussian interference
• Enables reliable classification at SNR levels below 0 dB
• Eliminates the need for explicit noise variance estimation

Raw cumulant values depend on signal power and phase rotation, which vary with channel conditions. Normalized cumulants and cumulant ratios (e.g., |C40|/|C42|²) create features invariant to amplitude scaling, while operations like |C41| remove phase dependence. This invariance means the downstream neural network does not waste capacity learning to compensate for nuisance parameters.

• |C40|/|C42|² is amplitude-independent
• Magnitude operations eliminate carrier phase offset sensitivity
• Reduces required training data by removing channel variability

A raw IQ sample block may contain thousands of complex values, but a cumulant-based feature vector typically consists of only 5–20 carefully selected statistics. This extreme dimensionality reduction provides a highly compressed, information-dense representation that enables smaller, faster neural networks with lower inference latency and reduced overfitting risk.

• Typical feature vector: 6–12 cumulant values and ratios
• Enables lightweight classifier architectures (2–3 hidden layers)
• Ideal for FPGA and edge deployment with limited memory

Cumulant features naturally support coarse-to-fine classification hierarchies. A single fourth-order cumulant can separate PSK from QAM modulations at the root node, while additional cumulant ratios refine the decision to specific orders (e.g., QPSK vs. 8-PSK, 16-QAM vs. 64-QAM). This structured approach reduces multi-class confusion and provides interpretable decision paths.

• Root node: C42 separates sub-Gaussian (PSK) from super-Gaussian (QAM)
• Subsequent nodes use C40, C60, C80 for order identification
• Enables rejection of unknown modulations at each hierarchy level

Unlike raw IQ-based deep learning classifiers that can be fooled by imperceptible waveform perturbations, cumulant-based features exhibit inherent adversarial robustness. Higher-order statistics aggregate information over many samples, making them insensitive to small, localized distortions. An attacker must inject significant energy to shift cumulant values, which is easily detectable.

• Sample aggregation provides natural smoothing against perturbations
• Cumulant estimators are low-pass statistical filters
• Shifts the attack burden from subtle to conspicuous interference

Cumulant features support recursive update equations that refine estimates with each new IQ sample without storing the entire batch. This enables continuous, streaming modulation classification where the feature vector evolves in real time, eliminating the latency of block-based processing and enabling instantaneous modulation change detection.

• Recursive moment updates: M_k(n) = M_k(n-1) + (x_n^k - M_k(n-1))/n
• Converts to cumulants via moment-to-cumulant formulas
• Enables sample-by-sample classification with sub-millisecond latency

A structured set of estimated cumulants and their ratios concatenated into a single input vector for a machine learning classifier. This vector serves as the bridge between classical signal processing and deep learning, providing a compact, interpretable representation that captures the distribution shape of the received signal. Typical vectors include C20, C21, C40, C41, C42, C60, C61, C62, and C63, along with key ratios like |C40|/|C42| to ensure robustness to phase and frequency offsets. The dimensionality is fixed regardless of observation length, making it ideal for lightweight neural network architectures.

A scale-invariant cumulant value obtained by dividing a higher-order cumulant by a power of the signal variance. This normalization ensures the classification feature is independent of the received signal amplitude, a critical property for real-world deployment where automatic gain control may vary. Common normalizations include dividing C40 by C21² to produce a unitless statistic. Without normalization, a classifier would erroneously associate signal power with modulation type, destroying generalization across varying link budgets and receiver gains.

A discriminative feature formed by dividing two different cumulant orders, such as |C40|/|C42| or |C63|²/|C42|³. These ratios create modulation fingerprints that are inherently robust to phase rotation, frequency offset, and timing errors. For example, the theoretical |C40|/|C42| value is 1.0 for QPSK, 0.0 for 16-QAM, and 0.68 for 64-QAM, providing clear decision boundaries. Ratios eliminate the need for precise synchronization before feature extraction, a significant advantage in non-cooperative or blind classification scenarios.

An empirical estimate of the theoretical cumulant computed from a finite block of received IQ samples. The estimation uses k-statistics or direct moment-to-cumulant conversion formulas applied to the sample moments. The variance of the estimate decreases with the square root of the number of samples, establishing a fundamental trade-off between observation length and classification accuracy. Key practical considerations include: - Bias correction for small sample sizes - Recursive updating for streaming applications - Outlier sensitivity requiring robust estimation techniques

A preprocessing step that uses the second-order cumulant matrix (covariance) to decorrelate multi-channel signal data before applying higher-order cumulant algorithms. This spatial whitening removes color from the data, ensuring that subsequent cumulant tensor decompositions operate on uncorrelated components. In the context of deep feature extraction, whitening normalizes the input distribution, accelerating neural network convergence and preventing certain features from dominating the loss function due to scale differences across antenna elements or receiver chains.

A monitoring process that tracks the statistical distribution of cumulant features over time to detect concept drift in the signal environment. By establishing a baseline cumulant profile for known modulations, the system can trigger model retraining or adaptation when the feature distribution shifts due to: - New transmitter hardware with different impairments - Changes in propagation environment - Introduction of previously unseen modulation formats This closes the loop between feature extraction and model lifecycle management, ensuring sustained classification accuracy in dynamic spectrum environments.