The dataset and code are for research purposes only. The results of this study are published in the following poster:
Haipeng Li, Chenggang Wang, Nirnimesh Ghose, Boyang Wang, "POSTER: Robust Deep-learning-based Radio Fingerprinting with Fine-Tuning," ACM Conference on Security and Privacy in Wireless and Mobile Networks (WiSec’21), June, 2021.
The pdf of this poster can be found below (open access by ACM).
https://dl.acm.org/doi/pdf/10.1145/3448300.3468253
Interview video from WiSec'21
https://ceas.mediaspace.kaltura.com/media/WiSec21_Poster_Haipeng_Interview/1_p1dho7lf
Demo Video 1 on running GNURadio code in this repository
https://ceas.mediaspace.kaltura.com/media/WiSec21_Poster_Haipeng_Video_GNURadio/1_bvujudvo
Demo Video 2 on running machine learning code in this repository
https://ceas.mediaspace.kaltura.com/media/WiSec21_Poster_Haipeng_Video_ML/1_c3booanp
In this project, we study radio fingerprinting. Radio fingerprinting is a way of distinguishing a transmitter from others of the same type by analyzing raw I/Q data in wireless communications. Each transmitter has a unique fingerprint due to hardware imperfections on radio-frequency circuitry. The hardware imperfections include I/Q imbalance, phase noise, frequency offset, and sampling off-set. Specifically, given I/Q data collected on a receiver side, a receiver decides which transmitter it is in radio fingerprinting.
Recent studies have shown that deep learning can obtain high accuracy in radio fingerprinting. However, as wireless channels keep changing in the real world, one primary challenge is that the deep neural networks for radio fingerprinting are not robust, where the training I/Q data are different from test I/Q data, especially in a cross-day scenario. We leverage fine-tuning to improve the robustness of deep-learning-based radio fingerprinting.
This repository contains GNU Radio source code for receivers and transmitters in our testbed and source code for deep neural networks and fine-tuning.
The GNU_radio directory includes source code for receivers and transmitters using HackRF Ones.
The rf directory includes source code for reading I/Q data and generating I/Q traces for a neural network.
The fine-tuning directory includes source code for three convolutional neural networks and fine-tuning.
The complex_rf directory includes source code for complex_value neural networks (no fine-tuning is supported at this point).
We built a testbed with 1 receiver and 5 transmitters in order to collect I/Q data. Each receiver/transmitter is a HackRF One (with ANT500 antenna) running with GNU Radio. We leverage the open-source GNU Radio code from https://github.com/bastibl/gr-ieee802-11 to establish WiFi transmissions (IEEE 802.11 a/g) with BPSK 1/2 modulation scheme between the receiver and transmitter. We captured the I/Q data at 2.45 GHz center frequency with 2MHz bandwidth and a 2MHz sampling rate.
We collected raw I/Q data before FFT, after FFT and after WiFi Frame Equalizer. We collected 3 transmissions from each transmitter. Each transmission lasts for 30 seconds. We collected data from two different days in an indoor environment. Only I/Q data after WiFi Frame Equalizer are used in our current experiments.
Our raw I/Q dataset can be found below (last modified: Feb 2024).
Note: the above link needs to be updated every 6 months due to certain settings of OneDrive. If you find the link is expired and you cannot access the data, please feel free to email us (boyang.wang@uc.edu). We will update the link as soon as we can. Thanks!
Note: the above dataset consists of raw I/Q data. To generate inputs for a neural network, you will need to use code in rf. You can choose your own parameters (e.g., the length of each I/Q trace, the number of I/Q traces per transmitter, etc.).
In our experiments, each I/Q trace has 288 consecutive I/Q samples. We randomly select 100,000 I/Q traces per transmitter. In other words, we have 500,000 I/Q traces overall for a neural network. We use 72% for training, 8% for validation, and 20% for testing.
If you have any questions, please feel free to contact us.
Haipeng Li (li2hp@mail.uc.edu), University of Cincinnati
Nirnimesh Ghose (nghose@unl.edu), University of Nebraska Lincoln
Boyang Wang (boyang.wang@uc.edu), University of Cincinnati
