HEEDFUL: Leveraging Sequential Transfer Learning for Robust WiFi Device Fingerprinting Amid Hardware Warm-Up Effects

Deep Learning-based RF fingerprinting approaches struggle to perform well in cross-domain scenarios, particularly during hardware warm-up. This often-overlooked vulnerability has been jeopardizing their reliability and their adoption in practical settings. To address this critical gap, in this work, we first dive deep into the anatomy of RF fingerprints, revealing insights into the temporal fingerprinting variations during and post hardware stabilization. Introducing HEEDFUL, a novel framework harnessing sequential transfer learning and targeted impairment estimation, we then address these challenges with remarkable consistency, eliminating blind spots even during challenging warm-up phases. Our evaluation showcases HEEDFUL’s efficacy, achieving remarkable classification accuracies of up to 96% during the initial device operation intervals-far surpassing traditional models. Furthermore, cross-day and cross-protocol assessments confirm HEEDFUL’s superiority, achieving and maintaining high accuracy during both the stable and initial warm-up phases when tested on WiFi signals. Additionally, we release WiFi type B and N RF fingerprint datasets that, for the first time, incorporate both the time-domain representation and real hardware impairments of the frames. This underscores the importance of leveraging hardware impairment data, enabling a deeper understanding of fingerprints and facilitating the development of more robust RF fingerprinting solutions.

💡 Research Summary

The paper addresses a critical yet under‑explored problem in radio‑frequency (RF) fingerprinting: the degradation of classification performance during the hardware warm‑up period that follows device power‑on. While deep‑learning‑based RF fingerprinting has shown promise for device authentication, prior work has largely attributed cross‑day performance drops to channel variations, overlooking the fact that even in static or wired environments the accuracy collapses within the first few minutes of operation. By collecting extensive time‑domain I/Q samples and eight hardware impairment metrics from 15 Pycom Wi‑Fi devices (802.11b and 802.11n) over a 30‑minute warm‑up window, the authors demonstrate that thermal effects cause oscillators, amplifiers, and power supplies to drift, leading to pronounced amplitude, phase, and frequency fluctuations. These effects stabilize only after roughly 12 minutes, creating a “warm‑up blind spot” for conventional classifiers.

To overcome this, the authors propose HEEDFUL, a two‑stage framework that couples targeted hardware‑impairment estimation with sequential transfer learning. In the first stage, a single‑input multi‑output convolutional neural network (CNN) is trained to predict the eight impairments directly from the raw I/Q waveform, thereby learning a representation that explicitly captures the physical variations of the radio front‑end. In the second stage, the pretrained impairment estimator is frozen and a device‑classification head is added; the model is then fine‑tuned on stable‑phase data (source task) and subsequently transferred to the warm‑up data (target task). This sequential transfer forces the classifier to rely on impairment‑invariant features learned from the stable regime while still being able to adapt to the transient characteristics of the warm‑up period.

Experiments show that HEEDFUL dramatically outperforms both a standard ResNet and a conventional CNN across all evaluated scenarios. During the initial 2‑minute warm‑up interval, HEEDFUL achieves 90 % accuracy (96 % at 6 minutes), whereas ResNet and CNN drop to 20 %–40 %. Cross‑day tests on a second‑day dataset retain 87 % accuracy for the 0‑6 minute window, compared to 20 %–43 % for the baselines. The framework also proves protocol‑agnostic: identical models trained on 802.11b data maintain high performance on 802.11n frames, confirming that the learned fingerprint resides in the hardware impairments rather than protocol‑specific signal structures. t‑SNE visualizations of the impairment space reveal clear clustering of devices, providing interpretability and confirming that the estimator captures device‑specific hardware signatures.

The contributions are fourfold: (1) a thorough empirical analysis of hardware warm‑up effects on RF fingerprinting, (2) an anatomical study of eight key impairments and their temporal evolution, (3) the HEEDFUL framework that integrates impairment estimation with sequential transfer learning to achieve robust warm‑up performance, and (4) the public release of a large‑scale Wi‑Fi RF fingerprint dataset (including raw I/Q and impairment labels) and the associated codebase, facilitating reproducibility and future research.

Limitations include the focus on static indoor environments; the approach’s resilience to mobility, multipath, or outdoor conditions remains to be validated. The set of eight impairments is fixed and may need extension for newer transceiver technologies. Moreover, the sequential transfer assumes that source (stable) and target (warm‑up) domains are not too divergent; large temporal gaps could still cause degradation, suggesting that online adaptation mechanisms would be a valuable extension.

In summary, HEEDFUL offers a principled, data‑driven solution to the warm‑up blind spot in RF fingerprinting, enabling reliable device authentication from the moment a device powers on. By explicitly modeling hardware impairments and leveraging transfer learning, the work sets a new benchmark for temporal robustness and opens avenues for applying similar techniques to other wireless standards such as 5G, BLE, and satellite links.