Scholarlink

<p>Acoustic side-channel attacks on mechanical keyboards pose a significant threat to data privacy, especially as remote work relies heavily on Voice over IP (VoIP) communications. However, it remains unclear if distinct mechanical keystroke signatures survive aggressive lossy audio compression. In this study, we propose ASCA-Net, a deep learning-based framework that synchronizes dynamic spectral gating, ResNet-34 classification, and semantic post-correction to reconstruct keystrokes from degraded audio. Evaluated on the Multi-Keyboard Acoustic (MKA) dataset, the system achieves a baseline raw acoustic classification accuracy of 99.87% on local recordings and 96.97% on heavily compressed VoIP streams. Furthermore, the pipeline operates with ultra-low latency (13.44&#xa0;ms) on lightweight hardware, processing 74.4 keys per second. While these results establish a robust baseline for channel-distortion resilience under controlled conditions, we acknowledge that unconstrained real-world deployments may introduce a 5–15% performance variance. Ultimately, this work proves that standard VoIP codecs do not inherently sanitize mechanical acoustic signatures, necessitating proactive defensive engineering in modern computing environments.</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

ASCA-Net: deep learning-based acoustic side-channel attack system for keystroke recognition across multiple platforms

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Karwan Rawf,
  • Brwa Hassan

摘要

Acoustic side-channel attacks on mechanical keyboards pose a significant threat to data privacy, especially as remote work relies heavily on Voice over IP (VoIP) communications. However, it remains unclear if distinct mechanical keystroke signatures survive aggressive lossy audio compression. In this study, we propose ASCA-Net, a deep learning-based framework that synchronizes dynamic spectral gating, ResNet-34 classification, and semantic post-correction to reconstruct keystrokes from degraded audio. Evaluated on the Multi-Keyboard Acoustic (MKA) dataset, the system achieves a baseline raw acoustic classification accuracy of 99.87% on local recordings and 96.97% on heavily compressed VoIP streams. Furthermore, the pipeline operates with ultra-low latency (13.44 ms) on lightweight hardware, processing 74.4 keys per second. While these results establish a robust baseline for channel-distortion resilience under controlled conditions, we acknowledge that unconstrained real-world deployments may introduce a 5–15% performance variance. Ultimately, this work proves that standard VoIP codecs do not inherently sanitize mechanical acoustic signatures, necessitating proactive defensive engineering in modern computing environments.