A decision-making framework for user authentication using keystroke dynamics

Increasingly sophisticated cyber attacks threaten critical infrastructures, requiring more trusted user authentication mechanisms. In this work, we propose a deep learning-based user authentication framework that combines keystroke dynamics with Siamese neural networks to differentiate legitimate users from impostors. A key challenge in this area is the variability in password lengths, which leads to different feature sizes and complicates model training. Our approach uses interpolation-based data fusion strategies to standardize the number of keystroke features, ensuring consistency across different datasets and password lengths. Through experiments on the fused CMU and KeyRecs datasets, we have evaluated the effectiveness of the proposed decision-making framework with adaptive threshold strategies. The threshold strategy determines how the final decision boundary is set with respect to the user's baseline typing behavior. We empirically evaluated the framework on fused data, achieving an equal error rate as low as 0.11-0.12, indicating strong efficacy in detecting insider threats. We show how the obtained Siamese neural network with triplet loss function can be used to distinguish genuine users from impostors even under different input conditions, contributing to more robust and scalable intrusion detection systems.