Keystroke Dynamics Authentication with MLP, CNN, and LSTM on a Fixed-Text Data Set

Keystroke dynamics is a biometric authentication factor with minimal detriment for user convenience during authentication. In this work, three different machine learning approaches are compared on an original multi-password data set previously published, consisting of passwords of varying length and complexity, with realistic user and attack data. Using Multi-Layer Perceptron, Convolutional Neural Network, and Long Short-Term Memory, we demonstrate that all three machine learning methods result in a high accuracy for user/attacker prediction making it a strong candidate to supplement a regular password as part of multi-factor authentication. We achieve an overall average accuracy of 97.4% with an average equal error rate of 2.6%. Our data set and the machine learning pipeline source code implemented are open-source for reproducibility and to facilitate further research.