Abstract
Wi-Fi monitoring has grown in popularity as a means of keeping track of a variety of activities. It has been used for multiple purposes ranging from health monitoring, activity monitoring, and much more. Wi-Fi is growing at a breakneck pace, owing to the increased use of wireless gadgets. The Wi-Fi signal is composed of Channel State Information (CSI) values used to record various human activities; they may also be used to determine the rate of breathing. The movement of a person's body part alters signal reflection, resulting in changes to the CSI. Similarly, breathing signals are sensed by an individual's chest moments.This research begins by extracting CSI from Wi-Fi frames using a Wi-Fi router and a Raspberry Pi. We then used the dataset to recognize human activities and monitor vital signs. Additionally, the datasets were utilized to develop deep neural network (DNN) models capable of classifying human activities (multi-class classification) and detecting apnea (binary-class classification). Additionally, we constructed a DNN model using a CSI dataset generated by Intel NIC 5300. We run three white-box attacks on the DNN models, which demonstrates that the performances of DNN models are significantly impacted.
Given that an adversary may readily mislead DNN models through the addition of well-designed perturbations, we suggest a solution to this problem by providing a defense against these three attacks. We retrained our DNN models by adding adversarial samples to the current dataset, strengthening the model's robustness, and significantly improving the models' performance following the attacks. The project is implemented using Raspberry PI and utilizing Google Colab, GPU, and CUDA cores.