Adversarial WiFi Sensing for Privacy Preservation of Human Behaviors

Abstract

Recent research on WiFi sensing focuses on the identification of a wide range of human behaviors with high recognition accuracy. However, these well-studied recognition techniques can cause privacy concerns due to the ubiquity of WiFi signal and comprehensive behavior information embedded therein. In this letter, we take the first attempt to develop an adversarial deep network architecture for human behavior preservation. Our goal is to make desirable private behaviors of a human being not recognizable by general classifiers, while the recognition of other ones remaining unaffected. To achieve this, we propose a novel loss function, using which our network is capable of intentionally modifying CSI data extracted from received WiFi signal, constraining the new classification results to match the adversarial requirements. Experimental results demonstrate that, with the proposed adversarial scheme, the recognition rate of the human behaviors needed to be protected can be significantly decreased, while still maintaining the accuracy of other ones desired to be identified. Our source codes are available at https://github.com/siwangzhou/WiFi-ADG.