Pluggable Subband-Decoupled Fusion and Unlabeled Capture Prototype Calibration: Bridging High-SNR Null-Subcarrier Regrowth and Low-SNR Coherent Preambles for Cross-Domain IoT RF Fingerprinting

Fengchong Yao, Jianbing Li, Qing Liu, Kefeng Song, Haitao Li, Song Wang, Kunpeng WangarXiv 2026 · arXiv

TL;DR

The authors introduce GSDF-CAPC, a modular framework for Radio Frequency Fingerprint Identification (RFFI) that bridges the performance gap between high-SNR and low-SNR regimes. By decoupling the processing of active and null subcarriers and using a blind SNR gate, it provides a "pluggable" upgrade for existing RFFI models, achieving state-of-the-art accuracy with zero retraining of the host model and minimal computational overhead.

What Problem Does This Paper Address?

Physical-layer Radio Frequency Fingerprint Identification (RFFI) is a zero-overhead method to authenticate IoT devices by identifying unique hardware impairments (like power amplifier nonlinearities) imprinted on their wireless signals. However, practical deployment faces a "Physical Crossover Paradox":

  1. In-Band Signals: Populated active subcarriers are resilient at low Signal-to-Noise Ratios (SNR) but degrade significantly under severe indoor multipath.
  2. Null-Subcarrier Regrowth: Distortions that "leak" into unpopulated null subcarriers are highly resistant to multipath at high SNR but disappear quadratically below the thermal noise floor.

Existing models usually focus on one regime or use "static fusion" that fails when SNR drops, often losing over 40 percentage points in accuracy. Furthermore, most RFFI models are rigid; they can't be easily updated to handle new environments (domain shift) or different hardware without expensive retraining.

Fig. 1. Physical crossover paradox and multi-baseline Pareto frontier: (a) on DataPort-WiFi (E2, E3 →E1, single-frame Direct), NS-Branch (95.7% clean), 5.3M McAFF (81.6% at 200 ns), and 4.1M SigMix (91.3% clean) illustrate the three-way regime trade-off: SigMix drops to 43.5% at 200 ns multipath, NS-Branch and McAFF collapse at ≤20 dB AWGN (4.7% and 27.4% at 20 dB), and unweighted Static-Sum (g ≡1) suffers −44.7 pp negative transfer, whereas pluggable GSDF gains +29.4 pp over SigMix and +16.0 pp over IB-Host at 200 ns (+48.6 pp over McAFF at 20 dB) with exact 0.00 pp low-SNR fallback; (b) on three-benchmark F1 (U-CAPC), zero-retrain GSDF improves all 8/8 backbones (72/72 wins), with GSDF-CAPC (93K, 57.50%, 1.82 ms CPU ONNX) surpassing 5.3M McAFF and 14.2M–15.9M 3-Seed McAFF.

Methods

The proposed GSDF-CAPC framework consists of two main innovations:

1. Pluggable Gated Subband-Decoupled Fusion (GSDF)

Instead of joint training, which leads to "gradient starvation" (where high-power in-band signals drown out the subtle null-bin cues), the authors decouple the two.

  • NS-Branch: A lightweight (46.3K parameter) encoder specifically for null-subcarrier regrowth.
  • Blind SNR Gate: A source-calibrated gate that uses the lag-64 periodicity of the IEEE 802.11a/g preamble to estimate SNR in real-time. It dynamically weights the fusion: at high SNR, it leverages both branches; at low SNR, it gracefully falls back to the host model to avoid injecting noise.
  • Pluggable Architecture: This branch can be attached to any frozen, pre-trained RFFI backbone (the "host") without retraining it.

2. Unlabeled Capture Prototype Calibration (UOT-CAPC)

To handle domain shifts (e.g., moving a device to a different room or hardware aging) at the edge, the authors use Unbalanced Optimal Transport (UOT). This aligns the "prototypes" (class centroids) from the source domain with unlabeled captures in the target domain. It specifically handles "traffic skew," where some IoT devices might send many more packets than others, preventing the calibration from collapsing into a single class.

Overview of the GSDF-CAPC architecture showing the decoupled branches and the blind SNR gate.

Datasets

The framework was evaluated on three major public IEEE 802.11a/g RFFI benchmarks:

  • DataPort-WiFi: A four-antenna cross-room transfer dataset.
  • UCLA WiSig: A multi-week, multi-receiver dataset with significant hardware drift.
  • ORBIT MobRFFI: A spatial-grid dataset for cross-receiver evaluation.

Key Insights and Ideas

  • The Fisher Information Cliff: The authors mathematically prove (Proposition 1) that null-subcarrier information decays quadratically ($O(\gamma^2)$) at low SNR, while active-band information scales linearly ($O(\gamma)$). This provides the theoretical foundation for why a dynamic gate is mandatory.
  • Decoupling is Essential: Jointly training active and null subcarriers is counterproductive due to the 20–35 dB power disparity. Decoupling allows the model to learn subtle intermodulation distortions without them being overshadowed by nominal symbols.
  • Zero-Retrain Generality: The framework's ability to improve 8 different backbones (from simple CNNs to complex attention models) without retraining them suggests a path toward "universal" RFFI upgrades.

Key Results

  • Universal Improvement: GSDF won all 72/72 paired runs across eight different backbones.
  • Accuracy Gains: On the DataPort-WiFi benchmark, it provided a +29.4 pp gain over SigMix and +48.6 pp over McAFF in high-multipath/low-SNR crossover regimes.
  • Edge-Ready Efficiency: The native "IB-Host + GSDF" model uses only ~93K parameters and achieves a single-core CPU latency of just 1.82 ms (via ONNX), making it highly suitable for deployment on resource-constrained edge access points.
  • Robustness: It maintained 0.00 pp loss at low SNR, proving the effectiveness of the exact host-fallback mechanism.

Limitations and Future Work

  • Sparse Bursts: While the model performs well on 5–10 packet bursts, its performance on extremely sparse, single-packet "fire-and-forget" transmissions remains a challenge for the calibration component.
  • Open-Set Rejection: The current work focuses on closed-set identification (identifying which known device sent a packet). Future work is needed to handle "unknown" devices (open-set rejection).
  • MIMO Support: The current framework is optimized for SISO/SIMO; extending these subband-decoupling principles to full MIMO preambles is a suggested next step.

Summary updated

Abstract

Zero-overhead physical-layer authentication at edge wireless access points (APs) is vital for securing resource-constrained Internet of Things (IoT) transceivers against credential cloning, yet cross-domain IEEE 802.11a/g radio frequency fingerprint identification (RFFI) faces an acute signal-to-noise ratio (SNR) crossover. Power-amplifier spectral regrowth on unpopulated null subcarriers resists indoor multipath at high SNR but decays quadratically below the thermal noise floor, whereas populated active in-band preambles survive low SNR yet degrade under severe multipath, causing static fusion to lose up to 44.7 percentage points (pp). To our knowledge, we propose the first RFFI framework that dynamically modulates active in-band waveforms and unpopulated null-subcarrier regrowth via a blind physical SNR gate: Pluggable Gated Subband-Decoupled Fusion with Unlabeled Capture-Aware Prototype Calibration (GSDF-CAPC). Avoiding the 20-35 dB subband gradient imbalance of joint training, GSDF pairs any frozen pretrained host backbone with a 46.3K null-subcarrier plug-in encoder (NS-Branch), fusing scale-standardized logits and gate-warped spherical embeddings under a source-calibrated lag-64 thermal-SNR gate with exact low-SNR fallback. Over unlabeled target AP captures, forward-only unbalanced optimal-transport calibration (UOT-CAPC) aligns shifted prototypes under packet-arrival skew and recovers parameter-free balanced Sinkhorn (U-CAPC) at equipartition. Across DataPort-WiFi, WiSig, and ORBIT MobRFFI, zero-retrain GSDF wins all 72/72 paired runs across eight backbones with 0.00 pp low-SNR loss, while edge IB-Host + GSDF (92.6K-93.3K active parameters, 1.82 ms single-core CPU latency) and timing-synchronized GSDF-STO / GSDF-Tri reach 57.50%-65.05% three-benchmark Macro-F1 (89.04%-100.00% over 5-10-packet IoT bursts).
BibTeX
@misc{yao2026pluggable,
  author = {Fengchong Yao and Jianbing Li and Qing Liu and Kefeng Song and Haitao Li and Song Wang and Kunpeng Wang},
  title = {{Pluggable Subband-Decoupled Fusion and Unlabeled Capture Prototype Calibration: Bridging High-SNR Null-Subcarrier Regrowth and Low-SNR Coherent Preambles for Cross-Domain IoT RF Fingerprinting}},
  year = {2026},
  eprint = {2611.22812},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22812},
}
Figure