A Novel Behavior-Based Tracking Attack for User Identification

Currently, people round the world daily use the Internet to access various services, such as, email and online shopping. However, the behavior-based tracking attacks have posed a considerable threat to users' privacy. Relying on characteristic patterns within the Internet activities, this attack can link a user's multiple sessions which are composed of a period of user's traffic. Once a user's personally identifiable information is disclosed in some session, the attacker can obtain the user's other network activities according to the linked sessions. In this paper, we investigate this behavior-based tracking attack and discuss the possible countermeasures. We preprocess the raw traffic data and then extract features ranging from lower layer network packets to high level application related traffic. Specifically, we focuses on four types of application-level traffic to infer users' habits, including HTTP, IM, Email, and P2P. A Multinomial Naive Bayes Classifier is employed to correlate users' sessions in distinct period. To evaluate the feasibility of our approach, we collect traffic in real-world environment to construct two distinct sizes of datasets. In the first dataset, we have 55 users' traffic during five weeks and the accuracy of our approach could reach 100%. To further illustrate the scalability of this approach, 509 users are selected from the second dataset in terms of the user's active degree. Finally, we can correctly correlate average 85.61% instances. Our extensive empirical experiments demonstrate the effectiveness and efficiency of our approach.