Using SVM for user profiling for autonomous smartphone authentication

While we have all been warned about viruses attacking our computers and hackers stealing our private information, very few of us realize the similar threat to our phones. With the number of smartphones in use growing each day, we now find ourselves to be a society equipped with devices packed with personal information and small enough to be easily stolen or misplaced. Millions of smartphone users are reporting unauthenticated behavior on their phones, yet many refuse to use passcode protection. Our goal is to use information gathered from the phone, specifically app usage statistics, in order to determine if a user is the actual owner of the smartphone. For this project, we created an app that could record a variety of information from smartphones and their sensors and make simple decisions about whether the person using the phone was the actual owner and lock itself accordingly. Because of time constraints, we looked at data sets from Glasgow Caledonian University and LiveLab at Rice University rather than collecting our own data. We used the LIBSVM library to create two-class SVM models for each of the 34 users in the LiveLab datasets. We then constructed testing datasets of both owner and non-owner data and tested the accuracy of the models. Accuracy rates for all 34 users were for the most part over 85%, and while false positive (identifying the owner as non-owner) rates were sometimes high, these false positive diagnoses would not compromise the security of the phone.