Phase Noise Fingerprint

The phase noise fingerprint manifests as a characteristic widening of the carrier signal in the frequency domain. Unlike ideal oscillators, real local oscillators exhibit short-term frequency instability that creates unique sideband noise skirts. These skirts follow a distinct power-law decay pattern (e.g., -30 dB/decade, -20 dB/decade) that varies between individual oscillators due to manufacturing tolerances.

• Close-in phase noise (1 Hz to 10 kHz offset) reveals flicker noise characteristics
• Far-out phase noise (100 kHz+ offset) exposes thermal noise floors
• The transition knee frequencies between noise regions are device-specific

Even identical oscillator models from the same production batch exhibit measurable phase noise differences due to microscopic manufacturing variations. These include:

• Crystal lattice defects in quartz oscillators affecting Q-factor
• Semiconductor doping inconsistencies in silicon-based VCOs
• Capacitor tolerance variations in PLL loop filters
• Thermal noise floor differences in active components

These physical variations create a unique phase noise signature that remains stable over the device's operational lifetime, making it a reliable biometric for RF emitter identification.

Allan variance (or Allan deviation) is the primary statistical tool for quantifying phase noise fingerprints in the time domain. It measures frequency stability as a function of averaging time, revealing distinct noise processes:

• White phase noise: σ² ∝ τ⁻¹ at short averaging times
• Flicker phase noise: σ² ∝ τ⁻¹ with different slope
• White frequency noise: σ² ∝ τ⁻¹/² (random walk FM)
• Flicker frequency noise: σ² reaches a floor
• Random walk frequency: σ² ∝ τ¹/² at long averaging times

The specific Allan deviation curve shape serves as a device-specific fingerprint for oscillator identification.

When a transmitter uses a Phase-Locked Loop (PLL) for frequency synthesis, the PLL components imprint additional identifiable artifacts on the phase noise fingerprint:

• Reference spur leakage at the phase detector comparison frequency
• Loop bandwidth corner frequency where VCO noise transitions to reference dominance
• Fractional-N spurs in synthesizers using delta-sigma modulation
• Charge pump mismatch signatures creating distinctive sideband asymmetry

These PLL-specific artifacts combine with the free-running oscillator noise to create a composite phase noise fingerprint unique to each transmitter's frequency synthesis chain.

Extracting phase noise fingerprints from received signals requires cross-correlation techniques to suppress additive channel noise and isolate the oscillator signature. The dual-channel measurement approach:

• Uses two independent receiver chains with a common reference
• Cross-correlates the demodulated phase error signals
• Suppresses uncorrelated receiver noise by √N averaging
• Reveals the transmitter's intrinsic phase noise below the receiver noise floor

This technique enables remote oscillator fingerprinting even through noisy channels, making it practical for over-the-air emitter identification at tactically relevant distances.

Phase noise fingerprints exhibit a temperature-dependent evolution that itself becomes an identifying characteristic. Each oscillator has a unique:

• Frequency-temperature curve with device-specific inflection points
• Phase noise floor temperature coefficient (dB/°C)
• Thermal hysteresis pattern during heating/cooling cycles
• Turn-on warmup transient lasting seconds to minutes

By modeling this thermal behavior, fingerprinting systems can maintain continuous authentication even as environmental conditions change, distinguishing genuine thermal drift from device substitution attempts.

The fundamental source of the phase noise fingerprint. Short-term frequency instability in the local oscillator (LO) causes random phase fluctuations. This is quantified by Allan variance in the time domain and single-sideband (SSB) phase noise (dBc/Hz) in the frequency domain. The unique phase-locked loop (PLL) dynamics and crystal oscillator characteristics imprint a device-specific spectral broadening on the carrier.

The physical mechanism by which phase noise degrades receiver performance and reveals the transmitter's fingerprint. In a receiver, a noisy LO mixes with a strong adjacent-channel interferer, smearing the interferer's spectrum into the desired channel. This raises the noise floor and is directly correlated to the LO's phase noise profile. Analyzing this effect helps isolate the LO contribution from other transmitter impairments.

Phase noise is one component of a transmitter's spectral regrowth and overall error vector magnitude (EVM). It must be distinguished from other impairments for robust fingerprinting:

• I/Q Imbalance: Creates a mirror image of the signal spectrum.
• Power Amplifier Non-Linearity: Causes spectral regrowth in adjacent channels via AM/AM and AM/PM distortion.
• Carrier Frequency Offset (CFO): A static shift, not a random fluctuation. Phase noise is unique because it is a dynamic, random process centered on the carrier.

Techniques like the bispectrum and cyclic autocorrelation are particularly effective at extracting phase noise signatures. Unlike the power spectral density, the bispectrum captures phase coupling information and is blind to Gaussian noise. This makes it robust for isolating the non-Gaussian, non-linear phase noise characteristics of a specific oscillator from additive white Gaussian noise (AWGN) in the channel.

Multipath fading can distort the phase noise fingerprint. Domain adversarial training is used to force a neural network to learn channel-invariant representations of the phase noise profile. The model is trained to identify the transmitter while simultaneously failing to predict the channel conditions, ensuring the fingerprint remains stable from an anechoic chamber to a dense urban environment.

Standard deep learning models process I and Q components as separate real-valued inputs, potentially losing the phase relationships critical to a phase noise fingerprint. Complex-valued neural networks (CVNNs) treat the signal as a single complex entity, using complex weights and activation functions. This preserves the magnitude-phase structure, allowing the model to learn more discriminative features from the subtle phase perturbations caused by the LO.