Breaking Anonymity of Social Network Accounts by Using Coordinated and Extensible Classifiers Based on Machine Learning

摘要

A method for de-anonymizing social network accounts is presented to clarify the privacy risks of such accounts as well as to deter their misuse such as by posting copyrighted, offensive, or bullying contents. In contrast to previous de-anonymization methods, which link accounts to other accounts, the presented method links accounts to resumes, which directly represent identities. The difficulty in using machine learning for de-anonymization, i.e. preparing positive examples of training data, is overcome by decomposing the learning problem into subproblems for which training data can be harvested from the Internet. Evaluation using 3 learning algorithms, 2 kinds of sentence features, 238 learned classifiers, 2 methods for fusing scores from the classifiers, and 30 volunteers' accounts and resumes demonstrated that the proposed method is effective. Because the training data are harvested from the Internet, the more information that is available on the Internet, the greater the effectiveness of the presented method.