A project I have headed for the past couple of years is
analyzing Graffiti input on Palm handhelds to see if we can reliably identify users
based on their writing style. I have done a lot of pre-processing, including analyzing
raw data for systematic and non-systematic errors and extracting features as well
as code for visualizing the data and the features. For detection I have applied various
statistical profiling techniques as well as ML algorithms, such as variations of
naive bayes and neural nets. We have obtained an average of over 80% hits with under
2% false alarm rates. This work, with SVM classifiers and near perfect accuracy rates with 6-12 character strings, has been accepted for publication at DSN 2007.