AIGC for RF Sensing: The Case of RFID-Based Human Activity Recognition

The performance of deep learning (DL) empowered wireless communications, networking, and sensing depends on the availability of sufficient high-quality radio frequency (RF) data, which is more difficult and expensive to collect than other types. To overcome this obstacle, we propose to harness the power of diffusion models on latent domains to generate hyper-realistic RF data for RF sensing. We develop a novel lightweight AIGC framework centered on latent domains, termed RFID-ACCLDM (Activity Class Conditional Latent Diffusion Model), to generate large quantities of RF data at low cost, conditioned on activity class labels. We demonstrate the high quality of RFID-ACCLDM generated data via the Frechet Inception Distance (FID) metric, along with a representative downstream task of human activity recognition (HAR). The model trained with synthesized data outperforms its counterpart trained by real data.