WiDistill: Distilling Large-scale Wi-Fi Datasets with Trajectory Matching

TL;DR

WiDistill introduces a novel dataset distillation method for Wi-Fi human activity recognition, reducing dataset size while maintaining high model performance and robustness across different environments.

Contribution

The paper presents WiDistill, a new approach that aligns parameter trajectories for effective Wi-Fi dataset distillation, enabling faster training and improved cross-network robustness.

Findings

WiDistill outperforms existing methods on multiple datasets.

It significantly reduces dataset size without sacrificing accuracy.

Models trained on distilled data perform comparably to those trained on full datasets.

Abstract

Wi-Fi based human activity recognition is a technology with immense potential in home automation, advanced caregiving, and enhanced security systems. It can distinguish human activity in environments with poor lighting and obstructions. However, most current Wi-Fi based human activity recognition methods are data-driven, leading to a continuous increase in the size of datasets. This results in a significant increase in the resources and time required to store and utilize these datasets. To address this issue, we propose WiDistill, a large-scale Wi-Fi datasets distillation method. WiDistill improves the distilled dataset by aligning the parameter trajectories of the distilled data with the recorded expert trajectories. WiDistill significantly reduces the need for the original large-scale Wi-Fi datasets and allows for faster training of models that approximate the performance of the original network, while also demonstrating robust performance in cross-network environments. Extensive experiments on the Widar3.0, XRF55, and MM-Fi datasets demonstrate that WiDistill outperforms other methods. The code can be found in https://github.com/the-sky001/WiDistill.