RF Fingerprinting And SEI

Specific Emitter Identification (SEI), also called RF fingerprinting, tries to identify which transmitter produced a signal from the analog hardware signature left in the received IQ. The framework’s DeviceFingerprint stage exists to generate synthetic data for this class of tasks.

Why These Impairments

The SEI literature converges on transmitter impairments such as CFO, SFO, IQ imbalance, phase noise, PA nonlinearity, and DAC quantization because they satisfy three useful constraints [1, 2]:

  1. Hardware-origin and difficult to clone. Values come from analog component tolerances, layout asymmetries, and manufacturing variation. Bit-similar SDRs can still differ measurably.

  2. Recoverable after propagation. Multipath and noise distort the signature, but CFO, IQ imbalance, PA statistics, and phase-noise spectral shape can survive enough to support identification.

  3. Stable on operational timescales. Real devices drift with temperature and aging, but within-session studies often assume stability over minutes to hours.

Other features in the literature, such as cyclostationary signatures, bispectrum, Hilbert-Huang transforms, wavelet decompositions, permutation entropy, differential constellation traces, and amplitude probability distributions, are derived from IQ after those impairments have already acted. The framework models the transmitter-side physics; downstream analysis can compute derived features from the synthesized IQ.

Time Stability

The framework default is fixed: one FingerprintInstance per device_id, applied identically across emissions from that device. This matches within-session SEI assumptions and gives a clean baseline.

That baseline is optimistic for cross-day evaluation. Real fingerprints drift with temperature, oscillator aging, and hardware state. To test cross-day robustness, generate two shards with the same device_ids but different temperature_c or time_offset values, then let the registry produce slightly perturbed realizations.

Evaluation Realities

Three results shape fingerprint-aware dataset design:

  1. Cross-receiver and cross-day accuracy collapse is the dominant failure mode. WiSig [3] showed that changing receiver or capture day significantly degrades classifier performance. Soltani et al. [4] showed that DNNs trained at one location and time perform poorly elsewhere. Cao et al. [5] argue that CNNs can entangle hardware fingerprints with reproducible environmental features, so the model may partly identify the path rather than the device.

  2. Channel-aware augmentation helps. Cömert et al. [6] compared noise injection, GAN augmentation, and 3GPP TDL/CDL channel profiling. TDL/CDL profiling improved transmitter recognition on unobserved data in their setup, which motivates pairing fingerprints with varied Sionna channel backends.

  3. Sim-plus-OTA methodology is standard. O’Shea et al. [7] established the pattern of simulating impairments and multipath, then validating against over-the-air SDR captures.

A dataset that varies fingerprints but holds the channel constant estimates an upper bound on discriminative power. A dataset that varies fingerprints, channels, receiver hardware, time, and location is closer to deployment.

Replay Resistance

SEI is not inherently replay-resistant. Recent work has shown that deep-learning classifiers can be defeated by SDR-based mimicry and replay attacks [5, 8].

Synthetic data should make hardening strategies easy:

  • Decoy emitters during training. Include devices labeled as unknown or adversarial so classifiers learn rejection behavior.

  • Hybrid signal-processing plus DL pipelines. Expose sync-word and preamble locations so downstream models can combine CFO estimation with learned features.

  • Distribution-shifted evaluation. Hold out devices, days, receivers, locations, and temperatures separately, not just random samples.

Open Problems

Active research directions include:

  • Channel robustness and scalability across propagation environments.

  • Domain adaptation across time and modulations.

  • Few-shot, semi-supervised, self-supervised, and contrastive SEI.

  • Adversarial and replay defense.

  • The effect of digital predistortion on identifiability.

  • Federated SEI across distributed receivers.

References

  1. Survey: RF fingerprinting / SEI methods, features, and classifiers. Peer-to-Peer Networking and Applications, 2024. https://link.springer.com/article/10.1007/s12083-024-01902-9

  2. Sankhe et al. ORACLE: Optimized Radio Classification through Convolutional Neural Networks. INFOCOM 2019. https://arxiv.org/abs/1812.01124

  3. Hanna et al. WiSig: A Large-Scale WiFi Signal Dataset for RF Fingerprinting. https://arxiv.org/abs/2112.15363

  4. Soltani et al. More Is Better: Data Augmentation for Channel-Resilient RF Fingerprinting. IEEE Communications Magazine, 2020. https://ieeexplore.ieee.org/abstract/document/9247526/

  5. Cao et al. On the Vulnerability of CNN-Based RF Fingerprinting to Replay Attacks. 2025. https://arxiv.org/html/2507.14109

  6. Cömert et al. Comparison of Augmentation Strategies for RF Fingerprinting Under Channel Mismatch. WiseML 2022. https://dl.acm.org/doi/abs/10.1145/3522783.3529518

  7. O’Shea et al. Over-the-Air Deep Learning Based Radio Signal Classification. 2017. https://arxiv.org/abs/1712.04578

  8. Replay Attacks Against SEI Classifiers. Discover IoT, 2024. https://link.springer.com/article/10.1007/s43926-024-00077-2