Data Diversity for a Channel-Resilient Training Database for Radio Frequency Fingerprint Identification

This paper proposes to improve the resilience of the Radio Frequency Fingerprint (RFF) identification to the propagation channel effects by massively diversifying the transmitted data. RFF refers to a unique signature resulting from signal distortions caused by the imperfections of the hardware transmission chains. In recent years, RFF identification has mainly relied on supervised Deep Learning (DL) techniques, which require an unbiased training database to avoid overfitting the propagation channel environment. This can be achieved during signal acquisition by acquiring many different channel realisations. However, designing such a database is very time-consuming, mainly because of the process of increasing the channel diversity. In this paper, we propose to change the paradigm of data diversity: instead of adding diversity to the channel, we choose to add diversity to the data. After a primary study to motivate the contribution, this paper presents, a proof of concept, which is carried out using model-based RF transmission databases, showing an improvement in data acquisition time of up to 16 times. A real database is then created using Software Defined Radios and the experimental results are compared with the simulated ones.