Short-Time Fourier Transform

The STFT applies a window function (e.g., Hamming, Hann, Blackman) to a short segment of the signal before computing the Fourier transform. This windowing isolates a quasi-stationary slice, mitigating spectral leakage caused by abrupt truncation. The choice of window directly impacts the trade-off between main lobe width and sidelobe suppression, determining the ability to resolve closely spaced frequency components in the RF fingerprint.

The STFT is governed by the Heisenberg-Gabor uncertainty principle, which states that time and frequency resolution cannot be simultaneously maximized. A narrow window provides excellent time localization but poor frequency resolution, ideal for capturing fast transients. A wide window yields fine frequency resolution but smears temporal events, making it suitable for analyzing stable steady-state oscillations. This trade-off is the central design parameter in RF fingerprint extraction.

The squared magnitude of the STFT, known as the spectrogram, visualizes the power spectral density of a signal as a function of time and frequency. For RF fingerprinting, the spectrogram reveals how phase noise, carrier frequency offset, and amplifier non-linearity manifest as distinct, time-varying spectral patterns. These patterns serve as the input feature map for deep learning models like Convolutional Neural Networks.

To avoid information loss at window boundaries, STFT computations typically employ overlapping windows (e.g., 50% or 75% overlap). The overlap-add method ensures that transient features occurring near the edges of one frame are captured in the next. This technique is critical for analyzing transient signal analysis events, such as the turn-on amplitude ramp of a transmitter, where a single window placement could miss the identifying signature.

The STFT transforms raw I/Q samples into a structured 2D representation where hardware impairments become visually and statistically separable. Key features extracted include:

• Spectral centroid drift over a burst
• Instantaneous frequency trajectories during symbol transitions
• Amplifier memory effect patterns visible as spectral spreading These features are often reduced via Principal Component Analysis before classification.

Unlike the Wavelet Scattering Transform, the STFT uses a fixed window size, resulting in a uniform time-frequency tiling. This makes it computationally efficient but less adaptive to signals with both fast transients and long steady-state components. The STFT excels at analyzing cyclostationary features where periodicities are stable, while wavelets provide superior multi-resolution analysis for non-stationary burst signals. The choice depends on the specific hardware impairment being targeted.

The visual output of the Short-Time Fourier Transform, plotting power spectral density against time. It reveals how a transmitter's carrier frequency offset and phase noise evolve during a burst. Key characteristics include:

• Resolution trade-off: A wide analysis window yields fine frequency resolution but smears transient events; a narrow window captures fast turn-on transients but blurs spectral lines.
• Feature extraction: Preamble-induced spectral ridges and amplifier non-linearity sidebands are directly visible as distinct patterns.

A mathematical weighting function applied to a signal segment before the Fourier transform to mitigate spectral leakage. The choice of window directly impacts fingerprint fidelity:

• Hann/Hamming: Good general-purpose windows balancing main-lobe width and side-lobe suppression for steady-state analysis.
• Kaiser: Offers a tunable parameter to precisely control the side-lobe level, useful for isolating weak local oscillator leakage near strong carriers.
• Rectangular: Equivalent to no windowing; high side-lobes can mask subtle impairments but preserve transient onset sharpness.

A fundamental limitation governed by the Gabor limit (uncertainty principle): the product of time resolution and frequency resolution cannot be arbitrarily small. For RF fingerprinting:

• Transient analysis demands high time resolution (short windows) to capture microsecond-duration turn-on amplitude ramps.
• Steady-state analysis requires high frequency resolution (long windows) to resolve closely spaced phase noise sidebands and I/Q imbalance spurs.
• Multi-resolution STFT applies different window lengths to the same signal to extract both classes of features simultaneously.

A method where consecutive STFT windows are overlapped (typically 50-75%) to avoid losing transient events that fall on window boundaries. This creates a smooth time-frequency representation without gaps. For device fingerprinting:

• Ensures the exact moment of a power amplifier turn-on transient is captured regardless of its alignment with the analysis grid.
• Increases computational load but prevents missed features in burst-mode communications like IoT protocols where the preamble is short and critical.

An alternative time-frequency representation that uses scaled and shifted basis functions (wavelets) instead of fixed-duration windows. Unlike the STFT's uniform resolution, wavelets provide:

• Multi-resolution analysis: Good frequency resolution at low frequencies and good time resolution at high frequencies, naturally matching many RF transient phenomena.
• Discrete Wavelet Transform (DWT): Efficiently decomposes a signal into approximation and detail coefficients, isolating hardware-induced discontinuities at specific scales.
• Often paired with STFT features in deep learning models to provide complementary views of emitter impairments.

An adaptive time-frequency method designed for non-linear and non-stationary signals, making it well-suited for analyzing distorted RF emissions. The process involves:

• Empirical Mode Decomposition (EMD): Data-driven sifting that extracts Intrinsic Mode Functions (IMFs) representing oscillatory modes inherent to the signal.
• Hilbert Spectral Analysis: Applied to each IMF to derive instantaneous frequency and amplitude, revealing how amplifier memory effects cause frequency deviations within a single pulse.
• Unlike the STFT, requires no predefined basis functions, allowing it to capture unexpected hardware behaviors.