KDAA: Korean Dubeolsik Acoustic Attack
An acoustic side-channel attack that reconstructs Korean text from Dubeolsik keystroke sounds, using an embedded CNN and a server-side jamo composition model.
Goal
Most acoustic side-channel attack (ASCA) research targets English QWERTY typing. KDAA extends this work to the Korean Dubeolsik layout, where recovering text also requires handling jamo composition, shifting syllable boundaries, compound jamo, Shift combinations, Korean/English mode switching, and Backspace.
Approach
KDAA is a two-stage pipeline:
- Acoustic front end: An ESP32-S3 converts keystroke audio into log-mel spectrograms and classifies them across 38 classes with a lightweight CNN. It then handles special keys and outputs top-N jamo candidates.
- Jamo Composition Model: A language model server turns the noisy jamo sequence into Korean text, with a Transformer reranking stage.
Evaluation
The CNN will be evaluated under three attack scenarios: close microphone, distant microphone, and background noise.
Status
Development and data collection tooling are in progress. Results will be added after evaluation. English input restoration is planned for v2.0.
Outputs 0
No publications or artifacts linked yet.