Adversarial Human Activity Recognition Using Wi-Fi CSI

Abstract

Human activity recognition has been used for various applications in Internet of Things (e.g., health monitoring, security, and sport-related monitoring). Wi-Fi channel state information (CSI) is widely used for activity recognition, where CSI can capture human activities that influence wireless channel. In this paper, we study the impact of adversarial attacks on deep neural network (DNN) based human activity recognition with Wi-Fi CSI. First, we discuss the system framework, where activity recognition can be considered as a classification problem and a specific DNN model is introduced. Then, we discuss adversarial attack problem for DNN-based human activity recognition and formulate three white-box attacks. In the experiment with a public Wi-Fi CSI dataset, our results show that the performances of DNN-based human activity classification are greatly influenced by three white-box adversarial attacks.