WiFi-based Intelligent Wireless Sensing for Privacy-Preserving Human Behavior Recognition under AIoT Architecture
Haoda WangPhD Candidate, Graduate School of Computer Science and Engineering, The University of Aizu, Aizuwakamatsu, Japan. d8242105@u-aizu.ac.jp0009-0009-8526-3988
Liang LinAssistant Professor, Department of Information Engineering, Luoding Polytechnic, Yunfu, China. 13016054668@163.com0000-0002-1161-5735
Huakun HuangAssociate Professor, School of Computer Science and Cyber Engineering, Guangzhou University, Guangdong, China. huanghuakun@gzhu.edu.cn0000-0003-2853-8892
Lingjun ZhaoAssociate Professor, Department of Network Engineering, School of Electronics and Information, Guangdong Polytechnic Normal University, Guangdong, China. atjonlyn@gmail.com0000-0003-2369-8862
Zhuotao LianPostdoctoral Fellow, Graduate School of Information Science and Electrical Engineering, Kyushu University, Fukuoka, Japan. zhuotaolian@ieee.org0000-0003-2938-6368
Chunhua SuSenior Associate Professor, Department of Computer Science and Engineering, Division of Computer Science, The University of Aizu, Aizuwakamatsu, Japan. chsu@u-aizu.ac.jp0000-0002-6461-9684
In the era of the Internet of Things (IoT) and AI-driven smart environments, human behavior recognition has emerged as a pivotal technology underpinning a broad spectrum of intelligent applications. However, achieving high recognition accuracy while preserving user privacy remains a critical challenge. To tackle this problem, this paper introduces a novel privacy-preserving method for WiFi-based human behavior recognition. It employs three-dimensional convolutional neural networks(3D-CNNs) enhanced with an attention-enabled autoencoder mechanism, called ThAN. The proposed approach also constructs three distinct classifiers activity, identity, and location that utilize 3D-CNNs for effective feature extraction from Channel State Information signals. Through rigorous experimentation with a real-world dataset, this study demonstrates that the method not only significantly safeguards privacy, achieving a high protection level, but also maintains a high recognition accuracy of 99%, substantiating its efficacy and applicability in real-world scenarios.