Can We Learn to Compress RF Signals?

The AI based on radio spectrum sensing is being deployed to detect and classify various interference sources and optimize spectrum allocation in next-generation cellular networks. Radio frequency (RF) spectrum sensing for AI must generate large quantities of RF data whose transport to the AI in the Cloud and at the Edge will incur significant bandwidth and latency costs. Can these quantities be compressed without affecting the utility of the cellular AI models? Our deep learned compression (DLC) model, named HQARF and based on learned vector quantization (VQ), compresses and reconstructs complex-valued samples of RF signals comprised of different modulation classes. In this paper we analyze how the salient (signal-processing) properties of the RF signal reconstructions by HQARF are modified due to incremental compression, and how this affects the accuracy of an AI model trained to infer the signal's modulation class. This analysis is also important as many advanced AI models developed for vision, such as Stable/VQ Diffusion, are trained on the quantized latent representation of the data. Understanding effects of learned RF quantization may help leverage those advanced models in the RF domain.