On the application of Supervised Time Series Forest to Radio Frequency Fingerprinting

Radio Frequency fingerprinting (RFF) is a method to authenticate wireless devices using their intrinsic physical features. It has been investigated by the research community in recent years, especially in combination with deep learning (DL) algorithms, which has demonstrated an excellent classification performance with the disadvantage of a high computing cost. In many studies where the original time domain representation of the signal is used, RFF can be considered a time series classification (TSC) problem and the various algorithms defined in the literature can be used to implement RFF. This possibility has been scarcely investigated in RFF literature where DL is usually preferred. A possible reason is also related to the elevated computing complexity of many TSC algorithms, which can be comparable to DL algorithms. In recent years, a new set of computationally efficient TSC algorithms has been proposed in literature, which can be suitable for a potential application in RFF. This paper investigates two aspects: 1) the application of the recently introduced Supervised Time Series Forest (STSF) algorithm to RFF and its comparison to CNN and other classifiers, and 2) the evaluation of STSF when combined in a hybrid approach of CNN with STSF (CNN-STSF), where the activation weights calculated by the CNN are used in input to the STSF in a similar way to hybrid CNN-ML algorithms presented in research literature. The proposed approach is applied to two public data sets showing that STSF has a very competitive performance in terms of execution time and classification accuracy.

BALDINI Gianmarco; 

2025-11-24

IEEE

JRC137259

https://ieeexplore.ieee.org/document/11104448,    https://publications.jrc.ec.europa.eu/repository/handle/JRC137259,   

10.1109/MeditCom64437.2025.11104448 (online),