Biometric Identification from Error Correction Behaviors Present in Keystroke Dynamics

This paper proposes two approaches for strengthening keystroke dynamics models- the reduction in size of usable key sets, and the analysis of relevant behaviors. Specifically, biometric identification using keystroke dynamics is analyzed when limited to deletion keystrokes and supplemented with extracted error-correction behaviors. To test this, participants typed with artificial spelling errors inserted, and a keylogger captured all deletion data. A neural network trained on this obtained an accuracy of $\mathbf{8 0 \%}$ with a loss of 0.61, indicating error-correction behaviors can be used for identification. This also demonstrates that a reduction in keysets can be effective for general keystroke dynamics.