Chaotic antenna arrays (CAAs) have been shown to produce enhanced and robust RF fingerprints, enabling more reliable device authentication using machine learning (ML) compared to traditional methods based on subtle hardware imperfections. In this paper, we present a novel CAA-specific multipath channel model to accurately capture the CAA’s phase errors across all propagation paths providing a basis for processing schemes that remove the channel effect, enabling extraction of the RF signature of the CAA. In addition, we provide a comprehensive experimental validation of CAA-based authentication under practical wireless channel conditions and include full details covering the design, fabrication, and characterization of custom CAA nodes, along with their integration within a software-defined radio (SDR) testbed for over-the-air measurements. This manuscript shows, for the first time, that training of the ML-based authenticator on the CAA RF fingerprints can be conducted under line-of-sight (LOS) conditions and effectively generalized to diverse scenarios, including non-line-of-sight (NLOS) environments, as long as a dominant propagation path exists. High authentication accuracy is consistently achieved when this key spatial channel characteristic is preserved. Accuracy exceeds 90% in LOS and reaches up to 87% in NLOS conditions with a single dominant reflected path. This study provides the first practical validation of spatially varying CAA fingerprints, underscoring their promise for secure and robust physical layer authentication across varied wireless conditions.