Twitter Fingerprints as Active Authenticators

Leveraging data drawn from the Web, or rather web analytics, has been used to gain business intelligence, increase sales, and optimize websites. Yet beyond the domain of ecommerce that web analytics is typically associated with, authentication based upon user interactions with the Web is also obtainable. Authentication is able to be achieved because just as individuals display unique mannerisms in everyday life, users interact with technology in unique manners. Leveraging these unique patterns, or "cognitive fingerprints", for security purposes can be referred to as active authentication. Active authentication stands to add extra security without added burden, as users are allowed the capability to simply interact with technology in their natural manner. Past research on active authentication has looked at areas such as mouse pattern movements, screen tough patterns on smartphones, and web browsing behavior. Our focus here is web browsing behavior. Specifically, we seek to extend past active authentication research done on Reddit. In this research, we examine the ability of Twitter-specific features to serve as authenticators, by examining the behavior of 50 random Twitter users. Through leveraging data mining and machine learning techniques, we conduct three levels of analysis: (1) we survey the ability of Twitter-specific behavioral features from a broad perspective to determine the feasibility Twitter fingerprints as a form of active authentication, (2) we compare aggregated and non-aggregated datasets to determine whether it is better to aggregate user behavior or look at posts individually, and (3) we examine whether certain features are more important for discrimination than others. The first level of analysis suggests that the posting behavior on Twitter follows the power law of human activity and that users can be uniquely identified with a fairly decent level of accuracy. Second, we find that aggregating the data significantly improves F-scores. Lastly, our examination suggests that there is not any specific feature that serves as more discriminative than others. Rather, what is discriminative for one user may not be for another user.