Bispectrum Analysis

The bispectrum of a Gaussian process is identically zero. This property makes bispectrum analysis exceptionally powerful for RF fingerprinting, as it theoretically eliminates additive white Gaussian noise (AWGN) from the extracted feature space. While real-world signals contain non-Gaussian interference, the method dramatically improves signal-to-noise ratio in low-SNR environments where second-order statistics fail.

Unlike the power spectrum, which discards all phase information, the bispectrum retains Fourier phase relationships. This is critical for device identification because hardware impairments—such as amplifier non-linearity and I/Q imbalance—manifest as quadratic phase coupling between frequency components. The bispectrum captures these phase correlations, which are unique to each transmitter's analog signal path.

The bispectrum quantifies the degree of quadratic phase coupling between frequency pairs (f1, f2) and their sum frequency (f1 + f2). This reveals non-linear interactions generated by:

• Power amplifier compression near saturation
• Mixer intermodulation products
• Local oscillator harmonics These coupling patterns form a distinctive, unclonable signature of the transmitter's analog front-end.

The bispectrum is shift-invariant—a time shift in the signal introduces only a linear phase term that cancels out in the bispectral magnitude. This property ensures that extracted fingerprints are robust to arbitrary signal start times and packet arrival jitter, eliminating the need for precise time synchronization during feature extraction.

Direct bispectrum estimation requires O(N²) operations for N frequency bins, making it computationally intensive for wideband signals. Practical implementations use:

• FFT-based indirect methods via the frequency-domain cumulant
• Segmented averaging (Bartlett or Welch methods) to reduce variance
• Radially integrated bispectrum (RIB) to compress 2D bispectral data into 1D features These optimizations are essential for real-time edge deployment.

Bispectral features serve as pre-processed inputs to convolutional neural networks for emitter identification. The 2D bispectral plane is treated as an image, where:

• Local bispectral peaks correspond to specific non-linear interactions
• CNN architectures learn hierarchical representations of coupling patterns
• Transfer learning from pre-trained vision models accelerates training This hybrid approach combines the noise-robustness of higher-order statistics with the pattern recognition power of deep learning.

The statistical building blocks of bispectrum analysis. While second-order statistics (autocorrelation) describe Gaussian processes, higher-order cumulants (third-order, fourth-order) capture non-Gaussian properties. The bispectrum is the Fourier transform of the third-order cumulant, making it sensitive to phase coupling and deviation from normality. Key properties:

• Zero for Gaussian noise (inherent noise suppression)
• Preserves phase information lost in power spectrum
• Detects quadratic phase coupling between frequency components

A non-linear phenomenon where two frequency components interact to generate a third component whose phase is the sum of the original phases. This occurs when a signal passes through non-linear hardware like a power amplifier. The bispectrum uniquely detects this coupling by measuring the bicoherence—a normalized bispectrum value indicating the degree of phase relationship. Transmitter-specific non-linearities create unique coupling patterns that serve as robust, unclonable fingerprints.

A defining advantage of bispectrum analysis for RF fingerprinting. Gaussian processes have a bispectrum of zero—theoretically eliminating additive white Gaussian noise (AWGN) that dominates wireless channels. This property enables:

• Extraction of subtle hardware signatures buried below the noise floor
• Robust operation in low signal-to-noise ratio (SNR) environments
• Cleaner feature representations for downstream neural network classifiers
• Reduced reliance on expensive pre-processing denoising stages

A normalized form of the bispectrum that quantifies the strength of quadratic phase coupling on a scale from 0 to 1. Bicoherence removes amplitude dependence, providing a pure measure of phase relationship. Estimation methods include:

• Direct FFT-based averaging (Welch's method adapted for bispectra)
• Indirect method using cumulant estimation and windowing
• Complex demodulation for non-stationary signals High bicoherence values at specific bifrequency pairs form distinctive patterns unique to each transmitter's non-linear hardware chain.

A complementary technique that exploits the periodic statistical properties of communication signals. While bispectrum analysis captures non-linear phase coupling, cyclostationary processing reveals hidden periodicities at cycle frequencies related to symbol rate, carrier offset, and modulation scheme. Combining both approaches yields:

• Bispectrum for non-Gaussian, non-linear signatures
• Cyclic spectral density for linear periodic features
• A multi-dimensional fingerprint robust to both noise and interference

The physical root cause that makes bispectrum analysis effective for device identification. Power amplifiers introduce AM/AM (amplitude-to-amplitude) and AM/PM (amplitude-to-phase) distortion when operating near saturation. These non-linear transfer functions:

• Generate harmonics and intermodulation products
• Create the quadratic phase coupling detected by the bispectrum
• Vary uniquely per device due to semiconductor manufacturing variances
• Exhibit memory effects that further enrich the bispectral signature