Device Signature Baseline

The baseline must encode a rich, multi-dimensional representation of the device's hardware impairments. This vector aggregates discriminative features extracted from various signal domains.

• I/Q Imbalance: Captures gain and phase mismatches in the modulator.
• Carrier Frequency Offset (CFO): Records the stable local oscillator inaccuracy.
• Phase Noise Profile: Characterizes the unique spectral skirt around the carrier.
• Power Amplifier Non-Linearity: Models AM-AM and AM-PM distortion curves. A single scalar metric like Error Vector Magnitude (EVM) is insufficient; the baseline requires a comprehensive feature vector to uniquely identify a device among thousands.

The integrity of the baseline depends entirely on the quality of the enrollment capture. This process must occur in a controlled, high signal-to-noise ratio (SNR) environment to isolate hardware impairments from channel effects.

• Static Channel: Enrollment is often performed via a direct cable connection or in an anechoic chamber to eliminate multipath distortion.
• Multiple Captures: The baseline is an average template computed from hundreds of transmissions to suppress thermal noise variance.
• Ground Truth Binding: The extracted features are cryptographically bound to a unique device identifier (e.g., a silicon serial number) to prevent identity swapping.

A baseline captured in one environment must remain valid when the device transmits in a completely different multipath environment. This requires the enrolled features to be inherently channel-invariant or processed through domain adaptation.

• Cyclostationary Features: Exploit the signal's periodic statistical properties, which are largely unaffected by stationary multipath fading.
• Higher-Order Statistics (HOS): Bispectrum analysis suppresses Gaussian noise and linear channel distortion, preserving the non-Gaussian signature of transmitter non-linearities.
• Contrastive Learning: Neural networks are trained to map signals from the same device to nearby points in an embedding space, regardless of the channel impulse response.

Hardware impairments are not perfectly static; they drift slowly over time due to temperature variation and component aging. A frozen baseline will eventually cause a high False Rejection Rate (FRR).

• Incremental Updates: The baseline is periodically updated with newly authenticated transmissions using an exponential moving average.
• Temperature Profiling: Separate baseline templates are stored for different operating temperature ranges, with interpolation between them.
• Anomaly Detection: A sudden, large deviation from the baseline is flagged as a potential spoofing attack, while a slow, gradual drift is accepted as natural aging.

For real-time authentication at scale, the baseline must be stored efficiently and compared against incoming signals with minimal latency.

• Dimensionality Reduction: Techniques like Principal Component Analysis (PCA) compress the raw feature vector into a compact, low-dimensional representation while preserving 99% of the discriminative variance.
• Vector Database Indexing: Baselines are stored as embeddings in a vector database, enabling fast Approximate Nearest Neighbor (ANN) search to match a probe signal against millions of enrolled devices in milliseconds.
• Edge Deployment: For IoT applications, the baseline for a local set of authorized devices is stored directly on an embedded controller or FPGA, eliminating cloud round-trip latency.

A practical baseline system must handle devices it has never seen before and enroll new devices with minimal data.

• Open Set Recognition: The baseline is stored with a calibrated distance threshold. A probe signal whose embedding distance exceeds this threshold is classified as an unknown or rogue emitter, triggering an alert.
• Few-Shot Enrollment: Using a Siamese Network architecture, a new device can be enrolled with as few as 5-10 transmission examples. The network learns a similarity function rather than memorizing specific classes, enabling instant onboarding without retraining the entire model.

The mathematical process of transforming a raw I/Q waveform into a compact, numerical representation that captures the most discriminative hardware impairment information. Key techniques include:

• Higher-Order Statistics (HOS): Calculating bispectrum or trispectrum features.
• Cyclostationary Analysis: Exploiting periodic statistical properties.
• Wavelet Decomposition: Isolating joint time-frequency details. This vector becomes the stored baseline for future comparison.

An algorithmic mechanism that continuously updates a device's fingerprint baseline to account for slow, natural variation. Temperature changes and component aging cause hardware impairments to drift over time. Without compensation, the False Rejection Rate (FRR) increases. Techniques involve:

• Moving average updates of the embedding.
• Adaptive Kalman filtering to track signature evolution.
• Periodic re-enrollment triggers when drift exceeds a threshold.

A neural network architecture that learns a similarity metric between pairs of signal samples. During enrollment, a probe signal is passed through the network alongside the stored baseline to generate a similarity score. This enables few-shot authentication, where a device can be verified with minimal training examples by directly comparing its live signature to the enrolled template in an embedding space.

The point on a Receiver Operating Characteristic (ROC) curve where the False Acceptance Rate (FAR) and False Rejection Rate (FRR) are equal. EER is the standard single metric for evaluating the accuracy of a baseline authentication system. A lower EER indicates a more reliable system that balances security (rejecting impostors) with usability (accepting legitimate devices).

A primary hardware impairment captured in a baseline, caused by mismatched gain or phase in the modulator's in-phase and quadrature branches. This creates a distinctive, elliptical distortion of the transmitted constellation. Because the mismatch is a function of analog component tolerances, it serves as a highly stable and unique identifier for the specific transmitter.