Enhanced Feature Selection and Generation for 802.11 User Identification

To provide user privacy, several anonymization techniques (e.g., pseudonyms applied to MAC addresses) have been proposed in 802.11 networks. However, recent research done by Pang et al. has demonstrated that pseudonyms are not adequate to protect user privacy. The key idea of Pang et al.'s method is to locate implicit identifiers (e.g., IP addresses and port numbers a user frequently visits), build user profiles based on these implicit identifiers in the training data sets, and then apply classification techniques to identify unlabeled (testing) users. Our method proposed in this paper partly focuses on building user profiles. Compared with the method proposed by Pang et al. in, we propose a novel approach to selecting and generating features, which is critical to build user profiles. The feature selection and generation procedure can be dynamically controlled through setting a few important parameters. We did a series of simulations using 9.27 GB SIGCOMM 2004 wireless data sets, and our simulation results demonstrate better classification rates compared with Pang et al.'s method.