Is it really you?: user identification via adaptive behavior fingerprinting.

The increased popularity of mobile devices widens opportunities for a user either to lose the device or to have the device stolen and compromised. At the same time, user interaction with a mobile device generates a unique set of features such as dialed numbers, timestamps of communication activities, contacted base stations, etc. This work proposes several methods to identify the user based on her communications history. Specifically, the proposed methods detect an abnormality based on the behavior fingerprint generated by a set of features from the network for each user session. We present an implementation of such methods that use features from real SMS, and voice call records from a major tier 1 cellular operator. This can potentially trigger a rapid reaction upon an unauthorized user gaining control of a lost or stolen terminal, preventing data compromise and device misuse. The proposed solution can also detect background malicious traffic originated by, for example, a malicious application running on the mobile device. Our experiments with annonymized data from 10,000 users, representing over 14 million SMS and voice call detail records, show that the proposed methods are scalable and can continuously identify millions of mobile users while preserving data privacy, and achieving low false positives and high misuse detection rates with low storage and computation overhead.