Bispectrum Fingerprinting

The bispectrum uniquely captures quadratic phase coupling (QPC) — the nonlinear interaction where two frequency components generate a third whose phase equals the sum of the parent phases. This reveals harmonic relationships introduced by power amplifier non-linearity and mixer imperfections that are invisible to the power spectrum. Unlike the power spectrum, which discards all phase information, the bispectrum preserves the phase relationships critical for discriminating between transmitters with identical spectral envelopes.

A defining mathematical property of the bispectrum is that it is identically zero for any Gaussian process. This provides a powerful theoretical advantage: additive white Gaussian noise (AWGN) — the dominant impairment in wireless channels — is asymptotically suppressed in the bispectral domain. The result is a feature representation with significantly higher signal-to-noise ratio (SNR) than raw I/Q samples or power spectral density estimates, enabling reliable fingerprint extraction even when the signal is buried below the noise floor.

The bispectrum is inherently shift-invariant to time-domain translations of the input signal. This property eliminates the need for precise time synchronization or preamble alignment during fingerprint extraction — a critical advantage in real-world intercept scenarios where signal start times are unknown. Combined with its insensitivity to linear phase shifts introduced by the channel, the bispectrum provides a representation that is robust to common signal acquisition imperfections without requiring complex pre-processing pipelines.

Most communication signals exhibit non-Gaussian statistics due to modulation constraints, amplifier saturation, and hardware impairments. The bispectrum quantifies the skewness (third-order cumulant spectrum) of the signal distribution in the frequency domain, capturing asymmetry and deviation from Gaussianity that directly reflects transmitter-specific hardware behavior. This makes it particularly effective for fingerprinting signals with constant-envelope modulations like GMSK and CPM, where amplitude-based features provide limited discrimination.

Modern SEI systems compute the bispectrum as a 2D feature map and feed it directly into convolutional neural networks (CNNs) for classification. The bispectral plane — with axes representing bifrequency coordinates — is treated as an image where texture patterns encode transmitter identity. Architectures like ResNet-50 and EfficientNet pre-trained on natural images are fine-tuned on bispectral maps, leveraging transfer learning to achieve high accuracy with limited training samples per emitter.

Direct bispectrum computation via the Brillinger-Rosenblatt estimator is O(N²) in the FFT length, creating a bottleneck for real-time applications. Practical implementations use radially integrated bispectrum (RIB) and axially integrated bispectrum (AIB) to reduce the 2D bispectral plane to 1D feature vectors while preserving discriminative information. Alternatively, diagonal slice bispectrum extraction computes only the bifrequency diagonal, reducing complexity to O(N log N) with minimal accuracy loss for many emitter classes.

The mathematical framework underpinning bispectrum analysis. HOS refers to statistical moments of order greater than two (variance). The third-order cumulant (skewness) and its Fourier transform, the bispectrum, capture phase coupling and non-Gaussianity that second-order statistics (power spectrum) completely miss. This makes HOS inherently immune to Gaussian noise, a critical advantage in low-SNR environments.

The unique signal characteristic that bispectrum analysis reveals. When a signal passes through a non-linear system (like a power amplifier), quadratic phase coupling occurs: frequency components at f1 and f2 interact to produce a component at f1+f2 whose phase is the sum of the original phases. The bispectrum detects this coupling, providing a fingerprint of non-linearity that the power spectrum cannot see.

A primary source of the hardware impairments that bispectrum fingerprinting exploits. When a transmitter's power amplifier operates near saturation, it introduces AM/AM and AM/PM distortion. These non-linear effects create characteristic phase coupling patterns in the bispectrum domain, forming a unique, device-specific signature that persists across different modulation schemes and data payloads.

A deep learning architecture that processes I/Q samples as complex numbers, preserving phase and magnitude relationships. When combined with bispectrum features, complex-valued networks can learn directly from the complex bispectrum without discarding phase information. This is critical because the bispectrum is a complex-valued function, and its phase contains the coupling information essential for fingerprinting.

A key design requirement for operational SEI systems. Multipath fading and Doppler shift can distort the bispectrum, reducing identification accuracy. Techniques to achieve channel robustness include:

• Domain adversarial training to learn channel-invariant features
• Signal conditioning (equalization, normalization) before bispectrum estimation
• Integrating the bispectrum along radial slices to create translation-invariant features