From the paper:
We built a prototype that can bootstrap the recognizer from about 10 minutes of English text typing ... For English text, the language constraints can be applied resulting in a 90-96% accuracy rate for characters and a 75-90% accuracy rate for words.There's some examples of the recovered text in the paper.
Remarkable and a little scary that, using nothing but a recording of typing and knowledge of patterns in the English language, someone can recover most of what was typed.
[via Bruce Schneier]
No comments:
Post a Comment