theowni/Keyboard-Typing-User-Recognition

Keyboard Typing User Recognition

Short description

This project aims for user recognition based on keyboard typing patterns using machine learning kNN algorithm, data based on flight and dwell time for keystrokes.

In future there could be more algorithms implemented just for fun with some clarification in docs.

Requirements

Linux/Mac/Windows
Python3.5
pip3

Usage

python3 app.py

Description

Data structure

Data was gathered via web page. There is about ~29200 records. Every analyzed record in database has that scheme:

INSERT INTO `user_typing_data` (`time`, `user_id`, `input0`, `IP`, `browser`) 
VALUES
('2013-11-20 07:25:12', 7, 'd_190_0_0 u_190_64 d_84_1008_1 u_84_1063
d_73_1776_2 u_73_1832 d_69_1872_3 u_69_1952 d_53_3128_4 u_53_3176
d_16_5256_5 d_82_5424_5 u_82_5464 u_16_5672 d_79_7303_6 u_79_7359
d_65_10167_7 u_65_10223 d_78_11420_8 u_78_11439 d_76_11679_9
u_76_11719 d_13_12156_10 u_13_12183 ', '81.219.51.76', 'Mozilla/5.0 (Windows
NT 5.1; rv:24.0) Gecko/20100101 Firefox/24.0'),

input0 field is list of items like that one:

(d/u)_keycode_time[_caret]

d/u - pressed/released action

keycode - js keycode

time - time elapsed from first pressed button (in ms)

caret - caret position

Data preprocessing

To extract only valuable numeric info about keyboard typing pattern, we process data to be list of items flight time (time button being pressed) and dwell time (time between pressing button) like this:

[60, 135, 60, 70, 39, 46, 79, 274, 80, 235, 59, 44, 81]

This trick makes two lists easy to compare using some metric for example Manhattan one. After that we can just use kNN algorithm which is widely described.

Results #1

Using kNN algorithm with k=5, effects below:

Test nr	Accuracy
1	0.867950
2	0.872765
3	0.842503
4	0.864512
5	0.861761
6	0.871389
7	0.867950
8	0.863824

For further extending algorithms there is multiple things which can be done to gain more accuracy, for example using IP address, user-agent or time when website was visited.

Results #2

Exteding distance using above parameters with some wages give quite nice accuracy improvement:

Test nr	Accuracy
1	0.936726
2	0.928473
3	0.940853
4	0.936726
5	0.927098
6	0.933975
7	0.931912
8	0.936039

It is possible to better adjust wages for parameters and gain more accuracy

theowni / Keyboard-Typing-User-Recognition