Sensitivity Analysis of RFML Applications

Abstract

Performance of radio frequency machine learning (RFML) models for classification tasks such as specific emitter identification (SEI) and automatic modulation classification (AMC) have improved greatly since their introduction, especially when measured against simulated data. When using captured RF data in a real environment, the performance of these RFML-based models is inconsistent when the propagation environment of the training data significantly differs from testing data. In this work, the correlations between measurable variations in propagation environment, ambient interference, amplifier compression, and overall classification performance are investigated and quantified. Parametric variations are ranked by impact to predict how well models trained in one environment can support operation in a dissimilar environment. Consistent with previous work, almost every factor studied was shown to impact classification performance in some way, with the effect of interference being particularly severe even at low levels.