These findings underscore two key strengths of our approach.
Foremost, the fact the system is able to perform at
84.5% accuracy (SD=18.9%) with a single training instance
from each user suggests the feature space we selected is
highly discriminative between users. Second, 0.5 seconds
appears to be a sweet spot, in particular, a nice balance between
classification accuracy and training duration. Whereas
an 8-second login sequence would significantly interrupt
interactive use, 500ms is sufficiently quick to be of minimal
distraction. Given that our current approach requires users
to login each time they want the system to differentiate
them, this interaction has to be extremely lightweight if it is
to be practical.