Software-Defined Radio (SDR)

Modern SDR platforms employ direct RF sampling to digitize signals at or near the antenna, eliminating analog down-conversion stages. This architecture uses high-speed analog-to-digital converters (ADCs) with sampling rates exceeding 3 GS/s to capture wide instantaneous bandwidths.

• Benefit: Preserves the raw, unfiltered signal including hardware impairments
• Example: The Ettus X440 captures up to 1.6 GHz of instantaneous bandwidth
• Impact on Fingerprinting: Captures the full spectral regrowth and phase noise sidebands that constitute a device's RF DNA

SDRs output in-phase and quadrature (IQ) samples as a complex baseband representation of the RF signal. This preserves both amplitude and phase information simultaneously, which is essential for analyzing hardware impairments.

• I Component: Represents the in-phase (cosine) projection of the signal
• Q Component: Represents the quadrature (sine) projection, 90 degrees offset
• Critical for SEI: I/Q imbalance itself is a key discriminating feature for emitter identification
• Data Rate: A 100 MS/s, 16-bit IQ stream generates 400 MB/s of data

High-performance SDRs integrate Field-Programmable Gate Arrays (FPGAs) for real-time, low-latency signal processing before data reaches the host CPU. This enables on-device operations critical for preprocessing RF fingerprints.

• Functions: Digital down-conversion (DDC), decimation, filtering, and packetization
• Benefit: Offloads computationally intensive tasks from the host, reducing data bottlenecks
• Fingerprinting Application: Real-time extraction of cyclostationary features or transient detection directly on the FPGA fabric
• Example: Xilinx RFSoC integrates ADCs, DACs, and FPGA logic on a single die

SDR platforms enable long-duration, wideband spectrum recording to disk for offline analysis and model training. This capability is fundamental for building large-scale, labeled datasets of RF emissions.

• Format: Raw IQ data stored in binary or SigMF (Signal Metadata Format) for interoperability
• Challenge: Sustained high-throughput storage requires RAID arrays or NVMe SSDs
• Application: Capturing hours of spectrum to train self-supervised learning models on unlabeled data
• Metadata: Timestamps, center frequency, and sample rate are stored alongside IQ data for reproducibility

The direct manipulation of in-phase and quadrature (IQ) data streams captured by an SDR. This includes correcting for I/Q imbalance, DC offset removal, and complex baseband representation. Proper IQ preprocessing is critical because hardware impairments in the SDR's own receiver chain can contaminate the transmitter's fingerprint.

The mathematical characterization of non-ideal behaviors in RF components that an SDR captures. Key impairments include:

• Power amplifier non-linearity: AM/AM and AM/PM distortion curves
• I/Q imbalance: Gain and phase mismatch between signal paths
• Oscillator phase noise: Spectral spreading of the carrier
• DC offset: Carrier leakage in direct-conversion receivers These impairments form the basis of a device's unique RF DNA.

A signal processing technique that exploits the periodic statistical properties of modulated signals. Unlike stationary noise, man-made communication signals exhibit cyclostationarity at symbol rates, carrier frequencies, and guard intervals. The spectral correlation function derived from SDR-captured IQ data provides robust, device-specific features resilient to stationary interference.

A transfer learning technique that mitigates the channel robustness problem in RF fingerprinting. When an SDR captures signals, the propagation environment (multipath, fading, Doppler) distorts the fingerprint. Domain adaptation aligns feature distributions between:

• Source domain: Training captures from one receiver/location
• Target domain: Operational captures from a different receiver/location This ensures the fingerprint, not the channel, drives classification.

A machine learning paradigm critical for operational SDR-based fingerprinting systems. Unlike closed-set classification, open-set recognition must:

• Correctly identify known emitters in the training set
• Detect and reject unknown or rogue devices never seen before
• Flag adversarial spoofing attempts This is essential for signals intelligence and physical layer security applications where new threats emerge constantly.

A deep learning architecture that learns a similarity metric between pairs of RF fingerprints. Rather than classifying emitters directly, a Siamese network maps IQ samples into an embedding space where:

• Signals from the same device cluster tightly together
• Signals from different devices are pushed apart This enables one-shot learning and clone detection by comparing a new signal to a stored reference fingerprint, even for emitters not in the training set.