Enhancing TypeFormer for Keyboard-Based Keystroke Authentication through Soft masking and Set-Margin Triplet Loss

This study proposes an enhanced training framework that extends TypeFormer for keyboard-based free-text keystroke authentication by introducing masking strategies and Set-Margin Triplet Loss (SM-TL). The masking mechanism effectively suppresses padding-related signals caused by variable-length free-text inputs, thereby improving model stability, while SM-TL learns class-level margin structures to produce more compact intra-class embeddings and clearer inter-class separations compared to the conventional Triplet Loss. Experiments conducted on the Aalto Desktop Dataset demonstrate that the proposed model consistently outperforms both TypeNet and the original TypeFormer in terms of EER. These results confirm that the enhanced TypeFormer offers strong practical applicability for desktop free-text authentication scenarios.