Abstract
Wi-Fi has been pivotal in changing the course of the world and technology ever since its inception. The number of applications that rely on Wi-Fi technology is staggeringly high. With the increase in the demand of wireless traffic data, using wireless data for various purposes has become common as they are easy to deploy and have high throughput. By exploiting the Channel State Information (CSI), we can obtain information about the surroundings, which gives us a peek into the future. This project provides a deep insight about collection of CSI data for gesture and activity recognition and creating a model to predict the type of gesture from the Wi-Fi CSI data. Although models nowadays are highly efficient, they do not possess a good security mechanism. In this project, along with the usage of CSI channels to gather information, we also attack the model created using adversarial attacks and reduce the efficiency of the model. Then we create defensive mechanisms for our model, which would make our model self-sustaining.
This project exposes the vulnerabilities of adversarial attacks and provides a solution to this problem by defending against these attacks. The project specially depicts the impact of adversarial attacks and how they can reduce the accuracy of any neural network-based models to be misguided into incorrect classification of data. The exposed vulnerabilities exhibit the ease with which information can be manipulated even without the knowledge of model parameters and dataset properties.