FlexiPrint: Adaptive Radio Fingerprinting for Length Variability

Wireless-enabled devices emit radio signals that contain unique characteristics embedded in the signal. This uniqueness forms a fingerprint that can be used in wireless communication systems for authentication at the physical layer and sensing applications. Deep learning is increasingly used to capture these unique characteristics for device classification. However, these approaches often do not consider signal length variation in real-world radio communication. This paper observes that a model trained on a fixed-length signal does not perform well when the length changes. To address this issue, we propose a novel approach called “FlexiPrint” that can maintain high classification accuracy regardless of the signal duration. We leverage the length of the signal to apply multiple inferences by analyzing different portions of the signal. We demonstrate that the proposed approaches maintain a high classification accuracy at ${9 8 \%}$ regardless of the signal length and without being trained on multiple lengths. Furthermore, FlexiPrint maintains high accuracy in noisy environments, outperforming the baseline approach, demonstrating robustness.