Scholarlink

<p>Acoustic side-channel attacks (ASCAs) pose a critical threat to digital security, enabling adversaries to decode keystrokes from typing sounds and compromise sensitive data like passwords. Despite advancements in keyboard security, modern devices remain vulnerable to these non-invasive attacks, particularly in public or remote settings (e.g., VoIP calls). This paper proposes an improved CoAtNet architecture to address this gap, combining convolutional layers for local feature extraction with transformer encoders to model long-range dependencies in acoustic signals. The model is trained on the Multi-Keyboard Acoustic (MKA) datasets—comprising six diverse platforms (HP, Lenovo, MSI, Mac, Messenger, Zoom)—and leverages Mel-spectrograms for robust feature representation, preserving discriminative acoustic details essential for distinguishing similar keystrokes. Experimental results demonstrate state-of-the-art performance in keystroke sound classification, achieving 99.8% accuracy, 99.81% precision, 99.8% recall, and 99.99% specificity, significantly outperforming prior methods—including the only comparable CoAtNet-based study reporting 95% accuracy on smartphone recordings and 93% on Zoom [Harrison, J., Toreini, E., Mehrnezhad, M.: A practical deep learning-based acoustic side channel attack on keyboards. in 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&amp;PW). 2023. IEEE]. This work establishes the first publicly reproducible, cross-platform benchmark for ASCA feasibility assessment, highlighting both the severity of acoustic vulnerabilities and the need for robust defensive strategies in modern human–computer interaction.</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

Improved CoAtNet for robust acoustic side-channel attack classification on keyboards

  • 期刊论文

  • 发表日期：

    2026年1月12日

查看全文
  • Karwan Mahdi Hama Rawf

摘要

Acoustic side-channel attacks (ASCAs) pose a critical threat to digital security, enabling adversaries to decode keystrokes from typing sounds and compromise sensitive data like passwords. Despite advancements in keyboard security, modern devices remain vulnerable to these non-invasive attacks, particularly in public or remote settings (e.g., VoIP calls). This paper proposes an improved CoAtNet architecture to address this gap, combining convolutional layers for local feature extraction with transformer encoders to model long-range dependencies in acoustic signals. The model is trained on the Multi-Keyboard Acoustic (MKA) datasets—comprising six diverse platforms (HP, Lenovo, MSI, Mac, Messenger, Zoom)—and leverages Mel-spectrograms for robust feature representation, preserving discriminative acoustic details essential for distinguishing similar keystrokes. Experimental results demonstrate state-of-the-art performance in keystroke sound classification, achieving 99.8% accuracy, 99.81% precision, 99.8% recall, and 99.99% specificity, significantly outperforming prior methods—including the only comparable CoAtNet-based study reporting 95% accuracy on smartphone recordings and 93% on Zoom [Harrison, J., Toreini, E., Mehrnezhad, M.: A practical deep learning-based acoustic side channel attack on keyboards. in 2023 IEEE European Symposium on Security and Privacy Workshops (EuroS&PW). 2023. IEEE]. This work establishes the first publicly reproducible, cross-platform benchmark for ASCA feasibility assessment, highlighting both the severity of acoustic vulnerabilities and the need for robust defensive strategies in modern human–computer interaction.