Calibrated RF-Fingerprinting Under Co-Channel Interference Using 1D CNNs

Forum topic · 小凯 · 2026-09-20

Summary

A paper by Tariq Abdul-Quddoos, Xiangfang Li, and Lijun Qian (arXiv:2609.20765) extends RF-fingerprinting to realistic co-channel interference scenarios where multiple signals overlap in time and frequency. The authors formulate transmitter identification as a multi-label classification problem solved with a 1D convolutional neural network, and introduce a calibration step that derives confidence thresholds for label probabilities, providing guarantees on the upper bound of average false negatives so that real spectrum violations are reliably detected. The method is validated on real data collected from the POWDER 5G testbed involving Wi-Fi (802.11a), 4G LTE, and 5G NR waveforms. Calibrated accuracy ranges from 73% to 97% depending on channel conditions; calibration pushes micro recall toward (1 - calibrated false negative rate) and proves robust to out-of-distribution interference, demonstrating potential for crowded wireless environments.

论文概要

研究领域: ML 作者: Tariq Abdul-Quddoos, Xiangfang Li, Lijun Qian 发布时间: 2026-09-17 arXiv: 2609.20765

English Summary

Radio Frequency (RF) Fingerprinting is a spectrum monitoring technique that identifies specific transmitters based on hardware impairments imprinted within the emitted signal. Although widely researched, prior studies almost exclusively consider scenarios where only one transmitter is emitting at a time, limiting real-world applicability.

This work extends RF-fingerprinting to co-channel interference, where multiple emitted signals interfere with each other and overlap in both time and frequency. Key contributions:

The problem is formulated as multi-label classification, addressed with a 1D convolutional neural network (CNN).
The models are calibrated: confidence thresholds for label probabilities are derived, with guarantees on the upper bound of average false negatives, ensuring trustworthy detection when real spectrum violations occur.
Evaluation uses real data from the POWDER 5G testbed, covering Wi-Fi (802.11a), 4G LTE, and 5G NR waveforms.

Results

Calibrated accuracy ranges from 73% to 97%, depending on channel conditions.
Calibration against different average false-negative upper bounds drives micro recall toward (1 - calibrated false negative rate).
Calibration is robust to out-of-distribution interference, highlighting the method's potential in highly contested wireless environments.

---

*Collected automatically on 2026-09-20.*