Acoustic side-channel attacks (ASCA) on keyboards exploit the sound of typing to infer keystrokes, potentially revealing sensitive information. These attacks have been a known security concern, but their practical effectiveness has been constrained by the limited scale of datasets used in prior studies.

To address this, the paper introduces DECKER, a method built on domain-invariant embeddings. The goal is to enable better cross-keyboard extraction and recognition, allowing models trained on one keyboard to work on others without extensive retraining.

The approach is presented as a step toward more robust ASCA, though the abstract does not provide specific performance numbers. The work highlights the ongoing arms race between acoustic side-channel threats and defensive measures.