Description

The sound emitted by keystrokes can potentially be used to perform a side channel attack for guessing the text being typed. This page runs a program that demonstrates such type of an attack.

First, type some text in the canvas below to train the prediction model. Don’t type super fast - it’s recommended to try a Hunt and Peck style of typing first to at least get some positive results. The program will try to predict keys which have been pressed at least 3 times.

When ready with typing the training data, click the Pedict button, wait for the model to train and continue typing outside the browser window. The program will continuously try to guess what is being typed by only analyzing the audio capture from the page.

Currently, the code is not very portable, so please read the requirements below and make sure you meet them. Also, this is a proof-of-concept and probably won’t work effectively in all environments. It does work pretty good on my setup though - iMac, Filco mechanical keyboard, built-in mic.

Requirements

  • Non-mobile browser supporting pthreads with WebAssembly (i.e. Chrome or Firefox Nightly)
  • Mechanical keyboard
  • Microphone (increase input level to max)


    • To enable WebAssembly pthreads and SharedArrayBuffer support, check this page.

    [Updated: 2 Dec 2018] Video: a working setup

    Demo