Revolutionizing User Authentication Exploiting Explainable AI and CTGAN-Based Keystroke Dynamics

Abstract

Due to the reliability and efficiency of keystroke dynamics, enterprises have adopted it widely in multi-factor authentication systems, effectively strengthening user authentication and thereby boosting the security of online and offline services. The existing works that detect imposter users suffer from performance and robustness degradation. Therefore, this article introduces a novel methodology to enhance user authentication and identify imposter users who attempt to have unauthorized access. We first use quantile transformation (QT) to mitigate outliers in the user's typing behavior that affects the authentication process and then employ conditional tabular generative adversarial networks (CTGAN) for data augmentation to learn the users' typing patterns better. Next, five accurate transfer learning models (VGG19, EfficientNetB0, Resnet50, MobileNetV2, and DenseNet121) are utilized for extracting effective features within the typing patterns, so our methodology can detect imposter users accurately and hence make precise decisions to enhance the user authentication process. Finally, we ensure transparency and trust in our user authentication methodology by incorporating explainable artificial intelligence (XAI), utilizing local interpretable model-agnostic explanations (LIME). Extensive experiments using a publicly available keystroke dynamics benchmark dataset from Carnegie Mellon University (CMU) showcase superior security performance and robustness using the proposed methodology compared to the state-of-the-art approaches.