YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Acoustic Keystroke Leakage Dataset (Multi-Microphone)
This dataset is designed for research in acoustic keystroke side-channel attacks, speech/typing synchronization, and keystroke dynamics. It features multi-channel audio recordings of typing tests captured under varying physical and digital configurations.
🚀 Data Gathering Framework
Data Collection Phases
The typing tests are structured into three distinct phases representing different cognitive and text-semantic profiles:
- Phase 1: No-Semantic text (syntactically cor-rect but semantically meaningless sentences generatedautomatically, minimizing linguistic predictability andemphasizing acoustic discrimination).
- Phase 2: Semantic text (coherent paragraphsin natural language).
- Phase 3: Spontaneous text (freely described images, capturing unconstrained typingbehavior, including pauses and corrections).
🎙️ Hardware & Microphone Setup
The dataset captures synchronized acoustic data from three distinct sources to simulate realistic physical and digital attack surfaces:
| Audio Source | Physical Location | Recording Computer | Characteristics & Limitations |
|---|---|---|---|
| Microphone 1 | Back of the keyboard | Victim's Computer | Direct, close-range acoustic signature of keys. Tied tightly to local logging. |
| Microphone 2 | Left of the keyboard | Attacker's Computer | Continuous, uncut audio stream. Note: Absolute keystroke timings based on local Unix timestamps cannot be fully synchronized with this channel due to independent clocks between the attacker and victim hardware. |
| Zoom_Audio | Digital Feed (Mic 1) | Attacker/Remote | The audio from Microphone 1 passed through a Zoom call. The platform's software noise-suppression filter was explicitly set to Minimal (Low). |
📊 Dataset Statistics
The table below breaks down the number of split utterances across dataset partitions (Train, Validation, Test), recording channels, and typing phases.
| Subset | Train | Validation | Test | Total |
|---|---|---|---|---|
| All | 3876 | 579 | 1157 | 5612 |
| Mic1 | 1216 | 185 | 371 | 1772 |
| Mic2 | 1332 | 197 | 370 | 1899 |
| VOIP (Zoom) | 1328 | 197 | 416 | 1941 |
| Phase 1 (Non-Sem) | 1364 | 209 | 381 | 1954 |
| Phase 2 (Semantic) | 1645 | 283 | 457 | 2385 |
| Phase 3 (Spontan) | 867 | 87 | 319 | 1273 |
📁 Dataset Structure
The processed dataset is organized into three parallel directories containing split speech/typing segments:
Clean_Dataset_Sentences/
├── wavs/ # Audio segments in their native sample rates (.wav format)
├── keys/ # Character alignment data (.txt format)
└── timestamps/ # Fine-grained temporal alignment metadata (.json format)
1. wavs/
Contains the segmented audio files cropped around typing events. Files preserve their original capture sample rates (no downsampling applied) to retain high-frequency acoustic characteristics of specific switches.
2. keys/
Text files containing the string or explicit ID of each pressed key matching the audio timeline. For normal phases, this contains the text sequence. For split/punctuated phases, elements include normalized key identifiers (e.g., tracking explicit events like "period", "question", or "exclam").
3. timestamps/
JSON files containing an ordered list of floating-point values representing the relative or absolute timestamps (in seconds) for each individual pressed key within that segment. This allows for token-level acoustic mapping.
file name format
[audioSource]_[PhaseID]_[SessionID]_[SentenceID].[extension]
[audioSource]: The recording channel source. Maps to either Mic1, Mic2, or Zoom_Audio.
[PhaseID]: The typing test phase corresponding to the text type (1 for Non-Semantic, 2 for Semantic, 3 for Spontaneous).
[SessionID]: A unique numerical identifier representing a specific participant and recording session.
[SentenceID]: The sequential index of the split sentence or text block within that session.
- Downloads last month
- 167