Secure and Robust User Authentication Using Transfer Learning and CTGAN-based Keystroke Dynamics

Due to keystroke dynamics's reliability and effi-ciency, it has been widely used in multi-factor authentication schemes by enterprises for robust user authentication, hence enhancing online service security. Therefore, in this paper, we propose a novel methodology designed not only to enhance user authentication but also to detect imposter users who are trying to get unauthorized access. Our methodology uses transfer learning and conditional tabular generative adversarial networks (CTGAN)-based keystroke dynamics, and consists of a four-step process; mitigating outliers that hinder the model's performance through utilizing quantile transformation (QT), employing data augmentation through CTGAN, converting CSV data into 3D images, and concatenating features from three accurate transfer learning models (VGG16, ResNet50, and DenseNet121) followed by dense layers to detecting im-poster users accurately and hence making precise decisions to enhance the user authentication process. We conducted extensive experiments to evaluate our methodology using the Carnegie Mellon University (CMU) keystroke dynamics benchmark dataset that contains real typing patterns, and the results demonstrate superior security performance and robustness using our methodology by achieving an accuracy of 99.99% and equal error rate (EER) of 1 %, thereby, outperforming the state-of-the-art.