Multi-Model authentication using keystroke dynamics for Smartphones

By the year 2018 it is predicted that 84% of the total world’s population would own a smartphone device. Intruders that gain access to someone’s phone can easily obtain personal and sensitive data about them, such as emails and messages, or even impersonate the rightful owner. Although an intruder can easily get the owner’s smartphone’s password by looking over their shoulder, it would be very hard for them to match the owner’s typing signature. However if the owner’s typing signature is only based on one typing mode, such as standing still and typing with one hand, it would be very inconvenient for the owner to always stand still and type with one hand as otherwise they’d likely be marked as intruders and get frustrated over time. This research addresses exactly this. We found that models covering different typing modes and activities such as standing or moving and typing one handed or two handed, can be achieved with Equal Error Rate values as low as 0.44%. The best results were achieved using a Least Squares SVM classifier with RBF kernel. We also explore and discuss any differences in classifying between typing activities, Digraphs vs Trigraphs, feature importance, misclassification analyses and full typing sessions vs sentence by sentence based classification.