i sensed it was you: authenticating mobile users with...

I Sensed It Was You: Authenticating Mobile Userswith Sensor-enhanced Keystroke Dynamics

Cristiano Giuffrida Kamil Majdanik Mauro Conti Herbert Bos

VU University Amsterdam

11th Conference on Detection of Intrusions andMalware and Vulnerability Assessment

Egham, UKJuly 10-11, 2014

1 / 18I Sensed It Was You: Authenticating Mobile Users with Sensor-enhanced Keystroke Dynamics Cristiano Giuffrida

The Blossom of the Mobile Computing Era

Mobile Data

Presentations / briefing notes.

Address book information.

Personal photos, movies, and email.

Personal health, salary, and benefits information.

Access credentials for networks and applications.

Credit card and e-banking information.

Mobile Threats

Mobile Authentication

Password/PIN/Pattern-based authentication.

" Simple and widespread.

% No continuous authentication.

% Prone to guessing attacks (not mobile specific).

% Prone to smudge [WOOT’10] and shoulder-surfing [CCS’13] attacks.

Biometric authentication.

" A viable option for many mobile users.

" Amenable to continuous authentication.

" Several existing mechanisms: gaits, gestures, keystroke dynamics.

% Poor accuracy (> 5%EER) or prone to statistical attacks [CCS’13].

WWW: What We Want

High-accuracy biometric authentication for mobile devices.

Robustness against human attacks.

Robustness against statistical attacks.

Static authentication capabilities.

Continuous authentication capabilities.

Robustness against uncontrolled settings.

WWW: What We Want

" High-accuracy biometric authentication for mobile devices.

" Robustness against human attacks.

- Robustness against statistical attacks.

" Static authentication capabilities.

- Continuous authentication capabilities.

- Robustness against uncontrolled settings.

Sensor-enhanced Keystroke Dynamics

Soft keyboard on a mobile device

Scenario: User typing ’HELLO’

User 1 KeyDowns User 1 KeyUps

Keystroke dynamics

User 1 User 1 KeyDowns User 1 KeyUps

Sensor-enhanced keystroke dynamics

2l 3l 4l 5l 6l 7l 8l 9l 0l1l

Ql Wl Ul Il OlEl Rl Tl Yl Pl

Al Sl Dl Fl Gl Hl Jl Kl Ll Dell

Zl Xl Cl Vl Bl Nl MlShiftl

Sensor

samples

Keystroke

events

Feature extractionmodule

Trainingmodule

Detectionmodule

Gathering Keystroke Events

Modified Android keyboard intercepts and records keystroke events.

Records only events of interest (i.e., alphanumeric characters).

KD time: Timestamp associated to key-down events.

KU time: Timestamp associated to key-up events.

Gathering Sensor Data

Gyroscope Accelerometer

Relies on the Android API to record sensor values while typing.

Can sample sensor values at a high frequency (e.g., 17 Hz).

Gyroscope: measures device orientation on the 3 axes.

Accelerometer: measures device acceleration on the 3 axes.

Feature Extraction

Traditional keystroke dynamics

Feature Extraction

0.5-graph

1-graph

1.5-graph

Complete word

KUKD KUKD KUKD

1 feature for each n-graph between KD/KU events.

Keystroke dynamics: time interval associated to each n-graph.

Sensor dynamics: statistical metrics associated to each n-graph.

Detection

Features gathered in a labeled vector and normalized.

Feature vectors used to train a binary classification algorithm.

Algorithms:Once-class Support Vector Machines (SVM).Naive Bayes.k-Nearest Neighbors (kNN).Mean algorithm.

Distance metrics:Euclidean.Euclidean normed.Manhattan.Manhattan scaled.Mahalanobis.

Experimental Setup

Fixed-text authentication system in a controlled setting.

2 predetermined passwords: “internet” and “satellite”.

20 test subjects, 40 (typo-free) password repetitions.

Samsung Nexus S with a soft landscape keyboard.

Trained detector for each user using leave-one-out cross-validation.

Measured FAR, FRR, EER and averaged results across users.

Factors considered: window size, algorithm, sampling frequency.

Accuracy vs. Window Size

10,00%

0.5-graph 0.5-graph &1.0-graph

1.5-graph &0.5-graph &1.0-graph

internet (min) satellite (min)

1.0-graph 1.5-graph 2.0-graph 2.5-graph 3.0-graph 3.5-graph 4.0-graph 4.5-graph whole word

internet (min) satellite (min)

Accuracy vs. Detection Algorithm

sensors sensors & timings

Accuracy vs. Sampling Frequency

0,03 0,09 0,17 0,34 0,85 1,70 3,40 4,25 5,67 8,50 11,33 12,75 13,60 15,30 16,15 16,66 16,83 16,92 16,97 17,00

Frequency (Hz)

Summary

Sensor-enhanced Keystroke Dynamics (SKD): A new biometricauthentication mechanism for mobile devices.

Unagi: A fixed-text authentication system based on SKD.

Key results:

Movement sensors are suitable for biometric authentication purposes.

Sensors can drastically enhance keystroke dynamics accuracy.

Effective even with short passwords and low sampling frequencies.

Future work:

Applicability to free-text authentication and uncontrolled settings.

Robustness against statistical attacks.

I Sensed It Was You: Authenticating Mobile Userswith Sensor-enhanced Keystroke Dynamics

Thank you!Any questions?

Cristiano Giuffrida, Kamil Majdanik, Mauro Conti, Herbert Bos{giuffrida,k.majdanik,mconti,herbertb}@cs.vu.nl

VU University Amsterdam

i sensed it was you: authenticating mobile users with...

Documents