SwinFi: a CSI Compression Method based on Swin Transformer for Wi-Fi Sensing

Wi-Fi sensing is a transformative approach that enables a large of applications through CSI analysis. The challenge lies in the high computational and communication costs with the increasing granularity of CSI data. In this letter, we propose SwinFi, a pioneering solution that compresses CSI at the edge into a succinct feature image and reconstructs at the cloud for further processing. SwinFi employs a Swin Transformer-based autoencoder-decoder architecture that ensures SOTA performance in both CSI reconstruction and sensing tasks. We utilize a dataset for PIR task and conduct extensive experiments to evaluate SwinFi. The results show that SwinFi achieves the reconstruction quality with the NMSE of -37.74dB and the classification accuracy of 95.3% at the same time.

Wi-Fi-based Personnel Identity Recognition: Addressing Dataset Imbalance with C-DDPMs

Wireless sensing technologies become increasingly prevalent due to the ubiquitous nature of wireless signals and their inherent privacy-friendly characteristics. Device-free personnel identity recognition, a prevalent application in wireless sensing, is susceptibly challenged by imbalanced channel state information (CSI) datasets. This letter proposes a novel method for CSI dataset augmentation that employs Conditional Denoising Diffusion Probabilistic Models (C-DDPMs) to generate additional samples that address class imbalance issues. The augmentation markedly improves classification accuracies on our homemade dataset, elevating all classes to above 94%.