EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

From Keystrokes to Semantics: Acoustic Side-Channel Attacks on Typed Text Using Deep Neural Networks and Large Language Models

Benjamin Quattrone, Youakim Badr

IEEE Access · 2026 · 14 · 91771-91789

DOI: 10.1109/access.2026.3697132 ↗

Kurz gefasst

Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Benjamin Quattrone und 1 weiteren Autor:innen. Die Arbeit nennt 31 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Acoustic Side-Channel Attacks exploit keyboard acoustics to infer typed input. Existing evaluations predominantly focus on keystroke or character accuracy, which is well suited for attacks requiring exact symbol recovery. However, such metrics are insufficient for assessing semantic information leakage in naturally occurring, unconstrained text. This work presents EchoStrike, an end-to-end ASCA pipeline that combines a deep neural acoustic classifier with multi-stage decoding, including language model refinement, to reconstruct long-form typed text from passive audio recordings. Using SKAID (Synchronized Keystroke and Acoustic Inference Dataset) to support realistic long-text evaluation, we report keystroke classification performance under strict cross-participant testing and quantify reconstruction quality using complementary syntactic and semantic metrics. Under this setting, the acoustic model achieves 76.34% keystroke accuracy on an unseen participant, and staged decoding improves long-text reconstruction, yielding a Sentence-BERT semantic similarity of 82.19% in a representative long-form evaluation. We further compare how signal processing choices, model architectures, and decoding strategies affect performance, and discuss implications for threat modeling where semantic comprehension, rather than exact character recovery, drives real-world risk.

Zitieren

APA

Quattrone, B., & Badr, Y. (2026). From Keystrokes to Semantics: Acoustic Side-Channel Attacks on Typed Text Using Deep Neural Networks and Large Language Models. IEEE Access, 14, 91771-91789. https://doi.org/10.1109/access.2026.3697132

Vancouver

Quattrone B, Badr Y. From Keystrokes to Semantics: Acoustic Side-Channel Attacks on Typed Text Using Deep Neural Networks and Large Language Models. IEEE Access. 2026;14:91771-91789. doi:10.1109/access.2026.3697132

BibTeX
@article{Quattrone2026,
  author = {Quattrone, Benjamin and Badr, Youakim},
  title = {From Keystrokes to Semantics: Acoustic Side-Channel Attacks on Typed Text Using Deep Neural Networks and Large Language Models},
  journal = {IEEE Access},
  year = {2026},
  volume = {14},
  pages = {91771--91789},
  doi = {10.1109/access.2026.3697132}
}
Kostenlos starten →