Indoor WLAN Personnel Intrusion Detection Using Transfer Learning-Aided Generative Adversarial Network with Light-Loaded Database

The Internet of Everything (IoE) provides a platform that allows devices to be remotely connected, sensed, and controlled across the network infrastructure. The smart home in the era of the IoE is born on the basis of the high integration of emerging communication technologies such as big data, sensors, and machine learning. In this paper, we focus on wireless detection technologies using smartphones and computers in smart homes. Among them, the indoor Wireless Local Area Network (WLAN) personnel intrusion detection technology based on the database construction has become one of the comprehensive detection technologies by advantages of the convenient accessibility of the WLAN signal and minimal hardware requirement. However, the considerable labor and time cost involved in the database construction affects the popularity and application of database-based intrusion detection systems. To cope with this problem, we propose a new indoor WLAN personnel intrusion detection approach with the reduced overhead of the database construction. Specifically, first of all, the offline database is extended by fake Received Signal Strength (RSS) data, which are generated by the Generative Adversarial Network (GAN) based supervised learning from actual labeled RSS data. Second, the difference between the extended database and online RSS data caused by the time-variant environment noise is reduced by minimizing the Maximum Mean Discrepancy (MMD) between marginal distributions of RSS data through the transfer learning. Finally, the intrusion detection is achieved by classifying online RSS data with classifiers trained from the extended database. Furthermore, experimental results show that the proposed approach can not only perform well in reducing the database overhead and the difference of data in source and target domains, which are corresponding to the same environment state but also detect environment states with satisfactory accuracy.