SEI Concept Drift

The gradual, physically-induced evolution of a transmitter's fingerprint due to component aging. Power amplifier transistors experience threshold voltage shifts, local oscillators exhibit increased phase noise, and capacitors degrade, collectively altering the I/Q imbalance and non-linearity signatures that SEI models rely on. This is a monotonic drift pattern—the fingerprint changes in a slow, directional manner rather than fluctuating randomly.

Changes in the physical or electromagnetic environment that alter a signal's presentation without the transmitter hardware itself changing. Key factors include:

• Temperature swings affecting oscillator stability
• Battery voltage droop in mobile devices altering amplifier bias points
• Multipath profile changes when a transmitter moves to a new location This drift is often reversible—the fingerprint returns to baseline when the environmental condition normalizes.

An abrupt fingerprint change triggered when a transmitter switches operational parameters. Examples include transmit power level changes, modulation rate switching, or antenna selection. While the device identity is unchanged, the SEI model may perceive a completely different emitter. Mitigation requires mode-aware training where the model learns to disentangle intentional configuration changes from device identity.

A critical distinction in SEI monitoring:

• Data drift (covariate shift): The input signal distribution P(X) changes, but the decision boundary P(Y|X) remains valid. Example: a new channel condition alters the received waveform, but the underlying fingerprint-to-identity mapping still holds.
• Concept drift: The conditional distribution P(Y|X) itself changes. The same fingerprint features now map to a different identity, requiring model retraining or adaptation. SEI systems must distinguish between these two failure modes to trigger the correct remediation.

Quantitative methods for detecting SEI concept drift before classification accuracy collapses:

• Embedding drift monitoring: Track the mean and covariance of penultimate-layer embeddings over time using a multivariate Gaussian model; trigger an alert when the Mahalanobis distance exceeds a threshold.
• Population Stability Index (PSI) applied to softmax confidence scores to detect shifts in prediction certainty.
• Open-set rejection rate tracking: A rising percentage of samples flagged as 'unknown' often signals drift before misclassification occurs.

Architectural approaches to maintain SEI accuracy under concept drift without full retraining:

• Online fine-tuning with a small buffer of recently authenticated samples, using a low learning rate to prevent catastrophic forgetting.
• Domain adversarial training to learn channel-invariant and aging-invariant feature representations.
• Ensemble of temporal models, where a new classifier is periodically trained on recent data and combined with legacy models via weighted voting, allowing the system to track gradual fingerprint evolution.