A nonlinear feature transformation-based multi-user classification algorithm for keystroke dynamics

Chinmay Sahu, Mahesh Banavar

Abstract:

Multi-user authentication from keystroke data is an
open problem that needs to be solved. When an online account
or a common desktop is shared among multiple users, there
is a need to determine who the current user is at log-in. In
this paper, a non-linear feature transformation-based multi-user
classification algorithm is proposed. Quantile transformation is
proposed to map raw keystroke features to a uniform distribution
to limit outliers. Then, dimensionality reduction techniques such
as PCA, Kernel-PCA, and t-SNE are applied to the transformed
features to project into reduced feature space. Unsupervised
clustering algorithms such as DBSCAN and GMM are applied
in the reduced feature space to identify the number of users
accessing the system. Using these results, a k-nearest neighbor
search algorithm is used, in conjunctions with labeled clusters
to classify users. The algorithm is validated using the CMU
keystroke benchmark dataset and the MobiKey touch dataset.
Once we identify the number of users, we can successfully classify
users with an accuracy of over 93 percent.