Where You Tap Matters: A Probe-and-Model Benchmark for Open-Set RF Fingerprinting

Source: arXiv:2607.21564 · Published 2026-07-23 · By Gabriele Oligeri, Savio Sciancalepore, Ingrid Huso, Fatima Al-Mousawi

TL;DR

This paper addresses the underexplored question of how the position within the receiver processing chain where IQ samples are collected affects Radio Frequency Fingerprint Identification (RFFI) in an open-set scenario. RFFI aims to identify transmitters by characteristics embedded in their physical signals rather than relying on higher-layer identifiers. The authors perform a controlled experimental measurement campaign using 12 Software Defined Radios (SDRs) transmitting BPSK signals connected by coaxial cable to a receiver chain implementing standard blocks: carrier recovery (costas loop), automatic gain control (AGC), pulse shaping (RRC filter), and symbol timing recovery. They collect data at five probe points along this chain. Using a common reconstruction-error based open-set classification method with image-based autoencoders, they show that the ability to distinguish enrolled devices from unauthorized transmitters varies markedly depending on at which probe point the data is taken. Timing recovery and, to a lesser extent, carrier recovery probes yield significantly better separation of samples from known vs unknown transmitters compared to earlier stages such as raw radio samples or post-AGC/pulse shaping outputs. To check the robustness of this result against model design, they generate several alternative autoencoder architectures using large language models (LLMs) and find the ranking by probe remains consistent, with negligible gains over the baseline shallow AE. Overall, the study highlights the importance of selecting the probe point within the receiver chain, which dominates the performance of open-set RFFI more than the neural architecture design choices.

Key findings

• RFFI performance strongly depends on the probe point within the receiver chain where IQ samples are collected; timing recovery (symbsync) and carrier recovery (costasloop) probes yield significantly better open-set discrimination.
• At the symbsync probe, 8 of 12 transmitters can be identified at true acceptance rate (TAR) = 0.9 while maintaining a false acceptance rate (FAR) below 0.09, compared to FAR > 0.1 at earlier probes for similar TAR.
• Receiver chain stages like AGC and RRC filtering reduce intra-transmitter variability but also homogenize device-specific features, harming RFFI discrimination.
• LLM-generated autoencoder designs with increased capacity, sparsity, denoising, or latent-structure regularization do not outperform the baseline shallow AE at the chosen operating point, and mostly increase training time.
• The compact convolutional AE (SuperGrok) slightly improves performance at the symbsync probe but offers no advantage at other probes.
• Open-set RFFI reconstruction-based methods rely on a delicate balance between suppressing noise/variation and preserving hardware-specific impairments; probe choice affects this balance more than architecture complexity.
• All experiments were performed on a 12-transmitter, 5-probe dataset collected via coaxial cable to isolate hardware features from wireless channel effects.

Threat model

The adversary attempts to impersonate or masquerade as an enrolled transmitter to gain unauthorized access. The system assumes the adversary cannot modify receiver chain processing or spoof physical-layer device fingerprints exactly. The model is trained only on enrolled transmitter samples from a single probe point; during inference, it must detect out-of-distribution signals from unknown transmitters based on reconstruction error. Channel effects are excluded, focusing purely on device-specific hardware impairments as the fingerprint source.

Methodology — deep read

The authors study open-set transmitter identification via Radio Frequency Fingerprinting (RFF) using reconstruction-error anomaly detection on IQ samples collected at different points of a BPSK receiver chain. The threat model is adversarial open-set detection, where the model is trained only with enrolled transmitter samples and must reject unknown transmitters; channel effects are excluded via direct cable connection to focus on hardware impairments.

Data was collected from 12 USRP B200-mini-i SDR transmitters connected via 50cm RF coax cable with a 20 dB attenuator to a USRP X410 receiver. A fixed BPSK transmitter and receiver chain were implemented in GNU Radio 3.8 with no higher-layer protocols, continuously sending a deterministic byte sequence mapped to BPSK. The TX and RX operated at 900 MHz sampling at 256 ksps with fixed gain and power settings chosen for stability. The receiver chain contains blocks for carrier recovery (Costas loop), automatic gain control (AGC), root raised cosine (RRC) filtering, and symbol timing recovery (Mueller and Muller TED).

IQ samples were tapped (probed) at five points: directly from radio source; after Costas loop; after AGC; after RRC filter; and after symbol synchronizer (symbsync). Each probe was saved as a dataset with about 153M samples for radio/costasloop/agc, 76M for rrc, and 38M for symbsync (due to decimation).

The core classifier pipeline converts 100,000-sample IQ chunks into 32x32 2D histograms of (|I|, Q) pairs with bin edges fixed from training data percentiles, applying log compression and 8-bit normalization. Images are vectorized and fed to an autoencoder trained unsupervised on enrolled transmitter samples from a given probe. Reconstruction error (MSE) per image is used as a score for open-set classification.

The baseline AE is a shallow image-based autoencoder trained with MATLAB's trainAutoencoder using a 64-unit bottleneck, logsig activations, L2 weight decay, and sparsity regularization.

The evaluation protocol fixes a transmitter and probe, training the AE on 70% training data from that setting and testing on held-out same-transmitter (in distribution, ID) and other-transmitters (out of distribution, OOD) samples. TAR and FAR are computed by varying the reconstruction error threshold to produce ROC curves. The key metric is the FAR at a fixed TAR=0.9.

To verify robustness, the authors use five popular LLMs (GPT5.2, Claude Sonnet 4.6, GLM5, Gemini 3.1 Pro, SuperGrok) via code generation prompts to create alternative AE architectures including convolutional autoencoders, smaller bottlenecks, denoising, latent center losses, and compare performance under the same preprocessing and scoring pipeline.

Training times and inference speed were recorded. Cross-probe generalization and enrollment were studied but details are limited in the current excerpt.

The dataset and code are planned for public release upon acceptance to facilitate reproducibility.

A concrete example: For transmitter #1 at the symbsync probe, 30 images were taken from unseen samples. The baseline AE reconstructed them, the MSE was computed per-image, and a threshold selected to yield 90% true acceptance. Reconstruction error distributions for ID and OOD samples showed minimal overlap at symbsync, with FAR below 0.1, demonstrating good open-set discrimination.

Overall, the methodology carefully isolates the effect of sampling point in the receiver chain on open-set RFFI performance, under a controlled hardware and channel environment and a fixed scoring pipeline.

Technical innovations

• Systematic experimental evaluation of open-set RFFI performance across five probe points along a standard BPSK receiver chain, revealing strong dependence on probe location.
• Use of LLM-assisted code generation to produce diverse alternative autoencoder architectures for RFFI, enabling an architectural stress test under fixed preprocessing and scoring.
• Proof that reconstruction error-based open-set RFFI performance is dominated by probe selection rather than autoencoder complexity or architecture enhancements.
• Demonstration that communication processing steps beneficial for demodulation (carrier and timing recovery) also enhance transmitter fingerprint separability by reducing intra-transmitter variability without removing device-specific residuals.

Datasets

• radio — 153 million IQ samples — collected at radio source output
• costasloop — 153 million IQ samples — after carrier recovery (Costas loop)
• agc — 153 million IQ samples — after automatic gain control
• rrc — 76 million IQ samples — after root raised cosine filter
• symbsync — 38 million IQ samples — after symbol timing recovery (symbol synchronizer)

Baselines vs proposed

• Baseline AE at symbsync probe: FAR = median ~0.03 at TAR=0.9
• Baseline AE at costasloop probe: FAR generally >0.05 but better than other probes
• Baseline AE at radio, agc, rrc probes: FAR >0.1 at TAR=0.9, indicating weaker separation
• SuperGrok CAE at symbsync probe: slightly improved FAR over baseline (~0.02 median), no improvement at other probes
• LLM-generated AEs (GPT5.2, Claude Sonnet 4.6, Gemini 3.1 Pro, GLM5): no consistent improvement over baseline; often higher FAR or training time

Figures from the paper

Figures are reproduced from the source paper for academic discussion. Original copyright: the paper authors. See arXiv:2607.21564.

Fig 1: Schematic representation of the GnuRadio receiver chain and points (probes) where the IQ samples are taken: radio,

Fig 2: IQ samples visualization example. We report 105 IQ samples taken from the probes radio (a), costasloop (b), agc

Fig 3: ¯ϵ as a function of all the possible pairs (Transmitter ID,

Fig 4: TAR as a function of the FAR while considering the

Fig 5: Performance (FAR given TAR = 0.9) as a function of

Fig 6 (page 4).

Fig 7 (page 4).

Fig 8 (page 4).

Limitations

• Dataset limited to 12 SDR transmitters connected by coaxial cable, excluding realistic wireless channel effects like fading or multipath.
• Open-set evaluation restricted to known 12 devices and fixed pre-processing pipeline; generalization to larger or more diverse device pools untested.
• Only BPSK modulation and a specific GNU Radio receiver chain tested; findings may not generalize to other modulations or chains.
• LLM-generated designs limited to autoencoder modifications under fixed preprocessing; no exploration of alternative scoring or classification methods.
• Performance improvements mostly benchmarked at fixed operating point TAR=0.9; other metrics or thresholds not thoroughly analyzed.
• Cross-probe transfer and enrollment dynamics not fully detailed or validated for operational deployment.

Open questions / follow-ons

• How do probe-dependent RFFI performances vary under real wireless channel conditions with fading, multipath, and noise?
• Can end-to-end learned approaches jointly optimizing probe selection, feature extraction, and scoring outperform fixed-probe reconstruction-error baselines?
• What is the impact of more diverse transmitter hardware, modulation types, and receiver architectures on the probe dependence phenomenon?
• Could multimodal approaches combining multiple probes or signal representations improve robustness and open-set identification?

Why it matters for bot defense

For bot-defense and CAPTCHA engineers implementing radio frequency fingerprinting for device authentication or anomaly detection, this work emphasizes the critical importance of selecting where along the receiver chain to collect IQ samples. Simply using raw or heavily processed signals may degrade fingerprint separability. Hence, practical RFFI-based systems should consider tapping signals after timing recovery or, at minimum, carrier recovery stages to retain discriminative device impairments for open-set classification. This insight guides data collection design to improve reliable device identification, reducing false acceptances of spoofed or unauthorized radios. The finding that more complex neural network architectures generated by LLMs yield limited benefits relative to probe selection also suggests that resource allocation is better spent on optimal data acquisition pipelines rather than solely on model complexity. Overall, the paper highlights that feature provenance within physical-layer pipelines is a fundamental design factor for successful RFFI in security applications.

Cite

@article{arxiv2607_21564,
  title={ Where You Tap Matters: A Probe-and-Model Benchmark for Open-Set RF Fingerprinting },
  author={ Gabriele Oligeri and Savio Sciancalepore and Ingrid Huso and Fatima Al-Mousawi },
  journal={arXiv preprint arXiv:2607.21564},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.21564}
}

Read the full paper

• arXiv:2607.21564 (abstract)
• arXiv:2607.21564 (PDF)