The Day-After-Tomorrow: On the Performance of Radio Fingerprinting over Time

TL;DR

This paper investigates the causes of the Day-After-Tomorrow effect in RF fingerprinting, revealing that radio power cycling significantly impacts performance and proposing pre-processing techniques to mitigate this issue.

Contribution

The study identifies radio power cycling as a key factor in the DAT effect and demonstrates how pre-processing can substantially improve fingerprinting accuracy in real-world scenarios.

Findings

Power cycling of radios causes significant performance degradation.

State-of-the-art RFF solutions double accuracy when radios are not power cycled.

Pre-processing of I-Q samples improves accuracy from 0.45 to 0.85 and reduces result variance.

Abstract

The performance of Radio Frequency (RF) Fingerprinting (RFF) techniques is negatively impacted when the training data is not temporally close to the testing data. This can limit the practical implementation of physical-layer authentication solutions. To circumvent this problem, current solutions involve collecting training and testing datasets at close time intervals -- this being detrimental to the real-life deployment of any physical-layer authentication solution. We refer to this issue as the Day-After-Tomorrow (DAT) effect, being widely attributed to the temporal variability of the wireless channel, which masks the physical-layer features of the transmitter, thus impairing the fingerprinting process. In this work, we investigate the DAT effect shedding light on its root causes. Our results refute previous knowledge by demonstrating that the DAT effect is not solely caused by the variability of the wireless channel. Instead, we prove that it is also due to the power cycling of the radios, i.e., the turning off and on of the radios between the collection of training and testing data. We show that state-of-the-art RFF solutions double their performance when the devices under test are not power cycled, i.e., the accuracy increases from about 0.5 to about 1 in a controlled scenario. Finally, we show how to mitigate the DAT effect in real-world scenarios, through pre-processing of the I-Q samples. Our experimental results show a significant improvement in accuracy, from approximately 0.45 to 0.85. Additionally, we reduce the variance of the results, making the overall performance more reliable.