Measurements and Analysis of Ambient Temperature on RF Fingerprinting Machine Learning Models

In this paper, we present a measurement analysis of machine learning-based Radio Frequency Fingerprinting (RFFP) techniques with respect to channel power and carrier frequency offset characteristics using custom Software-Defined Radio (SDR) hardware evaluated across different ambient temperature values. RFFP models have been demonstrated to classify transmitters based on features within the transmitted signal, although to validate these models when deployed in real-time applications, they must be invariant to changes in ambient temperature. Several papers have mentioned the potential impact of ambient temperature on RFFP classification performance, but to the best of the author’s knowledge, none have directly confirmed this assertion. In this work, we employ an Orthogonal Frequency Division Multiplexing (OFDM) waveform implemented using a custom SDR testbed to apply Analysis of Variance (ANOVA) and Design of Experiments (DOE) concepts to measure the effects of temperature, transmitted power, and center frequency on overthe-air machine learning data sets across a range of operating temperatures to assess performance. RFFP models based on these features are trained using over-the-air data sets at nominal temperatures, and the models are tested to see their performance when the radios operate at different temperatures. The results show that the RFFP classification models trained at 77°F(25°C) are not temperature-invariant at 63°F(17°C) and 95°F(35°C) and require new methods to maintain high classification accuracy.