Exploring Explainable AI Techniques for Radio Frequency Machine Learning

Deep learning models are increasingly being used to solve complex wireless radio frequency communications problems. These state-of-the-art machine learning models have demonstrated superior performance over traditional methods when signal and environmental parameters are unknown a priori. However, due to the complexity of the architecture and the number of parameters, deep learning models are difficult to interpret. This opacity can lead to difficulties during testing and a lack of trust by the user. Explainable artificial intelligence (XAI) techniques can provide estimates for the impact an input has on the output of a model. In this study, we apply a wide range of common attribution techniques, a subset of XAI that focuses on estimating the contribution of each input to an output of a model, to simple wireless communications problems over two different data modalities (raw IQ and spectrogram images) and show how estimates of attributions could be used for test and evaluation.