Abstract
Global Navigation Satellite Systems (GNSS) are utilized in many fields in everyday life. However, the number of radio-frequency interference (RFI) events, especially intentional ones, has been growing in recent years on the busy roads [1]. Such intentional interference is mainly jamming, meaning the transmission of radio signals at the GNSS frequency bands burying the real signal. The use of jamming devices is illegal in most countries, however, these jammers are easily bought from the internet and simple to use. Machine learning algorithms for radio-frequency signal classi-fication and fingerprinting exist. Each jammer has its own signal characteristics due to, for example, the non-linearities in the components arising from the manufacturing process [2]. However, the recording environment is not exactly the same at different times even at the same location. In order to reliably pair the signals with the transmitting devices, modeling the possible effects of the environment is important. To analyze these effects, we apply a baseline LSTM architecture to classify different jammers based on their transmitted signals. Also, Bayesian LSTM-based approach is introduced to model uncertainty in the predictions, that is, model's confidence. We train and test both networks with datasets acquired with different radio front-end devices in an anechoic chamber, first via transmitting only a jamming signal, and then together with GNSS signal repeated from an outdoor antenna. The understanding of environmental effects is the first step for developing robust fingerprinting methods. The main contribution of this paper is the analysis of the results that is further applicable for the next research step, namely the development of domain adaption algorithms to generalize the method over the effects.
