On the Impact of the Hardware Warm-Up Time on Deep Learning-Based RF Fingerprinting

TL;DR

This work investigates how hardware warm-up and stabilization affect deep learning–based RF fingerprinting, revealing that time-varying I/Q signal dynamics during warm-up can drastically degrade cross-time performance. By combining testbed experiments with simulations, the authors attribute these dynamics primarily to carrier frequency offset (CFO) and oscillator instability, demonstrating that stable data collected after stabilization yields robust fingerprinting, while unstable data collected during warm-up can mislead models. The paper provides quantitative analyses across wireless and wired channels, releasing a large multi-day WiFi 802.11b dataset and offering practical guidelines to improve robustness in security-critical wireless networks. The findings underscore the importance of accounting for hardware stabilization in RF fingerprinting deployments and datasets, particularly for cross-day and cross-setup scenarios.

Abstract

Deep learning-based RF fingerprinting offers great potential for improving the security robustness of various emerging wireless networks. Although much progress has been done in enhancing fingerprinting methods, the impact of device hardware stabilization and warm-up time on the achievable fingerprinting performances has not received adequate attention. As such, this paper focuses on addressing this gap by investigating and shedding light on what could go wrong if the hardware stabilization aspects are overlooked. Specifically, our experimental results show that when the deep learning models are trained with data samples captured after the hardware stabilizes but tested with data captured right after powering on the devices, the device classification accuracy drops below 37%. However, when both the training and testing data are captured after the stabilization period, the achievable average accuracy exceeds 99%, when the model is trained and tested on the same day, and achieves 88% and 96% when the model is trained on one day but tested on another day, for the wireless and wired scenarios, respectively. Additionally, in this work, we leverage simulation and testbed experimentation to explain the cause behind the I/Q signal behavior observed during the device hardware warm-up time that led to the RF fingerprinting performance degradation. Furthermore, we release a large WiFi dataset, containing both unstable (collected during the warm-up period) and stable (collected after the warm-up period) captures across multiple days. Our work contributes datasets, explanations, and guidelines to enhance the robustness of RF fingerprinting in securing emerging wireless networks.

Figures (13)

• Figure 1: I/Q signal behavior of Device A observed at different times during the device warm-up period.
• Figure 2: I/Q signal behavior of Device B observed at different times during the device warm-up period.
• Figure 3: I/Q signal behavior of Device C observed at different times during the device warm-up period.
• Figure 4: I/Q signal behavior of Device D observed at different times during the device warm-up period.
• Figure 5: The I component of the I/Q signal.