RF-CRATE: A White-Box Deep Learning Model for Wireless Sensing

A recent study has unveiled RF-CRATE, a white-box deep learning model designed for wireless sensing, which seeks to move away from traditional black-box methods in favor of mathematically interpretable designs. Published on arXiv under ID 2507.21799, this research tackles the deficiencies in physical and mathematical foundations present in existing Deep Wireless Sensing (DWS) models, which hampers their reliability and adaptability. RF-CRATE is based on the complex sparse rate reduction principle and employs the CR-Calculus framework to create a fully complex-valued transformer featuring interpretable self-attention and residual components. To address the challenge of limited labeled data, the authors propose subspace regularization, resulting in an average enhancement of 19.98%. The model has been tested on various RF modalities and human sensing tasks, such as activity, gait, and gesture recognition, marking a significant advancement towards white-box DWS for improved real-world application.