IQ Imbalance Drift

The core mechanism involves the slow temporal change in the relative amplitude (gain) and quadrature (phase) relationship between the I and Q branches. This is distinct from static IQ imbalance, which is a fixed offset. Drift manifests as a warping of the constellation diagram that evolves over hours, days, or months, directly impacting Error Vector Magnitude (EVM).

• Gain Drift: Causes the constellation to stretch or compress along one axis.
• Phase Drift: Causes the constellation to skew or rotate from its ideal square geometry.
• Combined Effect: Produces a non-orthogonal, asymmetrical constellation that degrades modulation quality.

The drift is fundamentally caused by the physical aging and environmental sensitivity of analog components in the modulator's direct-conversion architecture. The dominant root causes are thermal stress and semiconductor aging.

• Thermal Cycling: Repeated heating and cooling of the device physically stresses solder joints and alters the biasing conditions of transistors in the mixer and baseband amplifier stages.
• Hot Carrier Injection (HCI): A MOSFET aging mechanism where high-energy charge carriers get trapped in the gate oxide, permanently shifting the transistor's threshold voltage and altering gain characteristics.
• Passive Component Aging: Resistors and capacitors in the polyphase filter or quadrature hybrid can drift in value over time, directly changing the phase splitter's accuracy.

A critical characteristic for compensation is the distinction between reversible thermal effects and irreversible aging effects. A significant portion of observed IQ imbalance drift is a deterministic function of the component's junction temperature.

• Reversible Component: As the device cools down, the IQ imbalance parameters will partially return toward their previous state. This relationship is captured by the Temperature Coefficient of Impairment.
• Irreversible Component: The permanent offset that remains at a reference temperature represents true hardware aging. Effective drift compensation must model and separate these two superimposed effects to avoid over-correcting for a temporary thermal shift.

From a machine learning perspective, IQ imbalance drift causes a temporal covariate shift in the input feature space. A neural network trained on a device's initial constellation signature will experience degrading confidence as the drift pushes new samples away from the learned decision boundary.

• Concept Drift: The relationship between the raw IQ samples and the device identity changes. This violates the independent and identically distributed (i.i.d.) assumption of standard supervised learning.
• Manifold Warping: In a high-dimensional embedding space, the device's signature cluster slowly migrates. Without drift compensation, the system suffers from increasing false rejection rates (FRR) as the stored enrollment template becomes stale.

Drift is not uniform across a fleet of identical device models. Each physical unit possesses a unique drift trajectory due to microscopic manufacturing variances in its analog components. This is the same physical uniqueness that enables fingerprinting.

• Component Variability: Two identical transmitters from the same batch will age at different rates and in slightly different directions in the impairment space.
• Operational History: A device deployed in a hot outdoor enclosure will exhibit a fundamentally different drift profile than one in a climate-controlled data center. This necessitates per-device drift modeling rather than a one-size-fits-all population model.

Uncompensated IQ imbalance drift is a primary source of false rejection in long-term physical layer authentication systems. The system must distinguish a legitimate device whose signature has naturally drifted from an imposter attempting a replay attack.

• Security vs. Convenience Trade-off: Tightening the authentication threshold to detect subtle impersonation attacks increases sensitivity to normal drift, locking out valid users. Loosening it accommodates drift but creates a security vulnerability.
• Drift-Aware Similarity Metrics solve this by weighting features based on their known drift variance, effectively expanding the acceptance boundary in the direction of expected aging while keeping it tight against random imposters.

A recursive Bayesian algorithm used to estimate the true state of a drifting RF fingerprint by optimally combining a predictive aging model with noisy, real-time measurements. It operates in two steps: a prediction step that projects the IQ imbalance forward based on a state transition model, and an update step that corrects the prediction using a new observation. The Kalman gain dynamically weights the trust between the model and the measurement, making it ideal for tracking the slow, continuous variation of gain mismatch and phase error in quadrature modulators.

A statistical method for maintaining a device's reference fingerprint by applying a weighted average that gives higher importance to recent, authenticated transmissions while slowly forgetting older ones. The smoothing factor α controls the adaptation rate:

• High α: Rapid adaptation to drift, but more susceptible to noise
• Low α: Stable reference, but risks staleness This technique is computationally lightweight and effective for tracking gradual IQ imbalance drift without requiring complex predictive models.

A distance function modified to weight features based on their known drift rates, preventing false rejections due to normal aging. For IQ imbalance, the metric applies a temporal uncertainty covariance that expands the acceptance boundary along the expected drift trajectory. This ensures that a legitimate device whose gain mismatch has predictably shifted by 0.2 dB over six months is not falsely rejected, while an imposter with a random mismatch is still caught.

The creation of a mathematical model characterizing the precise, reversible relationship between a device's component temperature and its impairment values. IQ imbalance exhibits a temperature coefficient—a quantifiable change in gain and phase mismatch per degree Celsius. By measuring the device's temperature and applying a compensation curve, the reversible thermal component can be separated from the irreversible aging vector, ensuring that a hot device on day 100 still matches its temperature-normalized baseline from day one.

The Cumulative Sum control chart, a sequential analysis technique used to detect subtle but persistent shifts in the mean of a fingerprint feature. Unlike simple thresholding, CUSUM accumulates deviations over time:

• It triggers an alert when the cumulative sum exceeds a decision interval
• It detects small, sustained drifts in I/Q gain ratio or quadrature phase error long before they cause authentication failures This enables proactive signature refresh before the drift budget is exhausted.

A deep learning technique that trains a feature extractor to produce representations invariant to temporal domain shifts. A gradient reversal layer forces the network to learn IQ imbalance features that are discriminative for device identity but uninformative about the sample's age. The result: a fingerprint embedding where a device's representation from day one is directly comparable to its representation from day one hundred, effectively neutralizing the concept drift that would otherwise degrade classifier performance over time.