Starts at: March 31, 2014 2:00 PM
Ends at: 5:00 PM
Location: Silicon Valley Campus - Bldg 23, Rm 109
As mobile context-aware services gain mainstream popularity, and a smart phone accompanies its user throughout (nearly) all aspects of the user’s life, it is entrusted with an enormous amount of personal information, everything from context-information sensed by the phone to call-logs to social-media interactions to passwords. From this rich set of information it is possible to create models of the user’s (routine) behavior, and use these models to detect anomalous behavior. This will enable a variety of application domains such as device theft detection, improved authentication mechanisms, impersonation prevention, physical emergency detection, and remote elder-care monitoring. Existing work in this area mostly focuses on using a single sensor stream (typically GPS), greatly limiting the performance and coverage of these approaches. Another common trend in previous approaches is the use of a single monolithic model to learn user behavior, greatly increasing training data requirements.
This research presents a novel approach for modeling user behavior as a collection of models, each capturing a different aspect of the user behavior. A key component is a new ensemble learning approach, CobLE, for combining behavior aspect models to detect anomalous behavior. Evaluations performed on real-world datasets show this approach provides better anomaly detection performance with lower training data (i.e., boot-strapping time) requirements. Experimental results also show CobLE’s applicability outside of anomaly detection as a general ensemble learning approach, especially in situations with asynchronous data sources.