Typing Patterns for Authenticating and Identifying Users on Touchscreen Devices using Machine Learning

Authentication systems typically have their limitations and can be the target of attacks. The present research studies some of the contributions that the application of machine learning models can bring to the characterization of users by the way they interact with an application or service that works in touchscreen environments. The models will allow you to obtain a certain degree of confidence that the user who identified himself is true or not. The profile with which each user is identified is given by the characteristics of pressure, release, and time during the typing of the name and surname. For this, we use the dataset proposed by Morales containing 7200 samples collected from 300 users on the proposal to use the same data to validate our models to be adopted. With k-fold cross-validation, we train and evaluate three classification algorithms with an accuracy rate capable of helping in intrusion detection and identifying whether a user is trustworthy or an impostor. From the experiments carried out, the k-nearest neighbors and the Multilayer Perceptron Classifier outperformed the Support Vector Machine, achieving better classification robustness. The highest accuracy achieved was 82%. With these results, we observe that machine learning models can detect fraudulent users with relevant ranking performance. This allows us to conclude that the use of the username, password pair, and machine learning models as a combined strategy can allow maximizing security by discovering the user’s identity during authentication.