Inferring Behaviors via Encrypted Video Surveillance Traffic by Machine Learning
Xiaolong Liu, Jibao Wang, Ying Yang, Zigang Cao, Gang Xiong, Wei Xia · 2019
In order to protect the safety of people, smart cameras are widely used in many places of life: homes, offices, subways and many other ways. However, those smart cameras could lead to personal information leakage. Considering the role of smart cameras in peoples life, the information leakage may result in serious security problems. In this paper, we reveal the relationship between user behavior and encrypted video surveillance traffic. We show that different user behaviors have different traffic patterns. We propose some new features for encrypted video surveillance traffic, and we can infer the basic life behavior of users from these features. We exploit eight basic activities of daily living: watching TV, reading books, styling hair, open or close door, sweeping the ground, dressing, drinking and moving, which can cover the basic behavior of human beings. First, we show how attackers carry out an attack. We demonstrate it is easy for attackers to steal users' privacy by encrypted video surveillance traffic. Then, we extract the features of network traffic rate distribution and network traffic rate changes in the time domain and the frequency domain. Finally, we explore the validity of the features. And we utilize six typical machine learning algorithms to score these features. The experiment shows that we can identify the user's behavior with 94% accuracy.