StackFi: An Ensemble Learning-Based Model for WiFi Sensing Classification Tasks

The widespread adoption of WiFi devices and ubiquitous WiFi networks has fostered the development of the WiFi sensing domain. With the continuous progress in deep learning, a rising prevalence is observed in the application of deep learning techniques to the field of WiFi sensing. However, researchers often face the dilemma of choosing either convolutional neural networks, disregarding the temporal nature of Channel State Information data, or recurrent neural networks, overlooking the similarity between its subcarriers and the multiple channels in image data. We propose an ensemble learning-based model, StackFi, which combines two prevailing approaches in deep learning for processing Channel State Information (CSI) data. One approach analogizes CSI data to images, utilizing deep convolutional neural networks for feature extraction. The other treats CSI data as temporal data, employing recurrent neural networks for feature extraction. Our model extracts features from CSI data in WiFi signals, and these extracted features are then fed into a classifier. We gathered some commonly employed deep learning models in the field of WiFi sensing and contrasted them with StackFi using the public datasets UT-HAR and NTU-Fi HAR. Our model exhibited outstanding performance, achieving accuracy rates of 99.13% and 99.27% on these two public datasets, surpassing the performance of other models in the field.