Scholarlink

<p>With the growing urgency of information security needs, the traditional one-time authentication mechanism for identity verification has a significant flaw—it cannot provide continuous protection, and once breached, it leads to systemic risks. Keystroke authentication technology effectively addresses this issue by dynamically monitoring the entire session, capable of intercepting abnormal login attempts and potential threats in real time. However, existing methods, which rely on simple statistical features or single-scale temporal analysis, struggle to capture the dynamic patterns and multi-scale characteristics of keystroke behavior, resulting in insufficient accuracy in complex attack scenarios. To overcome these challenges, this paper proposes an innovative keystroke authentication model—TSFN. The model is centered around long short-term memory networks (LSTM) and Transformer encoders and ingeniously integrates our designed temporal gated convolutional block (TGBlock) and multi-scale dynamic weighted fusion block (MFBlock). These innovative designs enhance the model’s ability to process complex time-series data and extract features effectively. To validate the effectiveness of the TSFN model, we conducted experiments on both the Aalto desktop and mobile datasets. The results show that the model achieved ROC curve areas (AUC) of 99.17% and 99.06%, and the lowest equal error rates (EER) of 0.6% and 0.81% on the two datasets, respectively. Furthermore, the TSFN model achieves efficient real-time performance through parallel and distributed processing, meeting the demands for real-time identity verification on large-scale online platforms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

TSFN: a time-series fusion network for keystroke authentication

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Chang Wang,
  • Peiyu Li,
  • Yusong Lin

摘要

With the growing urgency of information security needs, the traditional one-time authentication mechanism for identity verification has a significant flaw—it cannot provide continuous protection, and once breached, it leads to systemic risks. Keystroke authentication technology effectively addresses this issue by dynamically monitoring the entire session, capable of intercepting abnormal login attempts and potential threats in real time. However, existing methods, which rely on simple statistical features or single-scale temporal analysis, struggle to capture the dynamic patterns and multi-scale characteristics of keystroke behavior, resulting in insufficient accuracy in complex attack scenarios. To overcome these challenges, this paper proposes an innovative keystroke authentication model—TSFN. The model is centered around long short-term memory networks (LSTM) and Transformer encoders and ingeniously integrates our designed temporal gated convolutional block (TGBlock) and multi-scale dynamic weighted fusion block (MFBlock). These innovative designs enhance the model’s ability to process complex time-series data and extract features effectively. To validate the effectiveness of the TSFN model, we conducted experiments on both the Aalto desktop and mobile datasets. The results show that the model achieved ROC curve areas (AUC) of 99.17% and 99.06%, and the lowest equal error rates (EER) of 0.6% and 0.81% on the two datasets, respectively. Furthermore, the TSFN model achieves efficient real-time performance through parallel and distributed processing, meeting the demands for real-time identity verification on large-scale online platforms.