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

TL;DR

SwinFi is a novel CSI compression method using Swin Transformer architecture that reduces computational costs while maintaining high accuracy in Wi-Fi sensing applications, enabling efficient edge-to-cloud processing.

Contribution

The paper introduces SwinFi, a pioneering CSI compression approach based on Swin Transformer, achieving state-of-the-art reconstruction and sensing performance with reduced data size.

Findings

Achieves NMSE of -37.74dB in CSI reconstruction

Attains 95.3% accuracy in PIR classification

Demonstrates effective edge-cloud CSI processing

Abstract

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.