Skip to content

Latest commit

 

History

History
66 lines (53 loc) · 2.57 KB

README.md

File metadata and controls

66 lines (53 loc) · 2.57 KB
Raw

The sound of typing: using Machine Learning to classify Keyboard Acoustic Emanations

BibTex citation:

@thesis{gkw2024soundoftyping,
    title={The sound of typing: using Machine Learning to classify Keyboard Acoustic Emanations},
    author={G{\'o}lski, Marcin and Kaszubski, Piotr and Woroch, Bart{\l}omiej},
    school = {Poznan University of Technology},
    address = {Piotrowo 3, 60-965, Pozna{\’n}, Poland},
    year={2024},
    type = {Bachelor's thesis}
}

This document will give a brief overview of how to recreate the results obtained in The sound of typing: using Machine Learning to classify Keyboard Acoustic Emanations.

Extracting keystroke peaks

Run ./recdata/wav_processing.py passing the names of the .wav files as arguments. The script will look for a matching .keys file within the same directory as the .wav file. To see additional options, run ./redata/wav_processing -h.

The files used for the experiments with peaks extracted this way are available in a separate repository, https://github.com/RoyalDonkey/put-kbd-thesis-datasets.

Combine the dataset into a single file

Run ./recdata/merge_files.py to aggregate the scattered .csv files created in the previous step into a single file representing the entire dataset. The parameters must be adjusted in the call to the merge_files() function. Note that the script expects a directory structure like this:

dataset_root/
    digits/
        ...
    letters/
        ...
    symbols/
        ...

Generate the results

Adjust the parameters within the source code of ./plot/create_results.py to match the desired datasets (DATASETS_LIST), peak types (wave_comb) and preprocessing techniques (preprocess_comb). Select the model to use in the section that identifies itself as appropriate (look for the comment: "you will probably need to adjust this part to particular models"). This will generate .json files with the accuracies the selected model achieved when being trained and tested on every pair of datasets found on DATASET_LIST. To do this for RNN, use create_results_rnn.py in the same way.

The plots can be recreated using other scripts found in the ./plot/ directory: best_acc.py, best_models_per_dataset.py, datasets_performance.py, model_comparison.py, and model_summary,py. Results and plots used for the thesis can be found in ./plot/results/ and ./plot/saved_plots/.

The data for the table showcasing top-3 rnn results was created with ./models/rnn/multkeys/ranked_network_results.py; the script requires an existing model instance to use to perform the predictions.