K Keystroke based User t k b d U Identification on Smart Phones - - PowerPoint PPT Presentation

RAID ‐ 2009

K t k b d U Keystroke‐based User Identification on Smart Phones Saira Zahid1, Muhammad Shahzad1, Syed Ali Khayam1,2, , , y y , Muddassar Farooq1

1 Next Generation Intelligent Networks Research Center 2 School of Electrical Engineering & Computer Sciences

g National University of Computer & Emerging Sciences Islamabad, Pakistan http://www.nexginrc.org g g p National University of Sciences & Technology Islamabad, Pakistan http://wisnet.seecs.edu.pk

How did the paper get accepted??? How did the paper get accepted???

• Extensive use of the words

–Smart Phones –Model Model

2 of 28

Motivation behind User Identification on Smart Phones

• Mobile computing devices combine three extremely

potent concepts

• computing

p p

• mobility
• miniaturization

y

3 of 28

Motivation (Contd )

SUMMARY FIGURE

M 2009

Motivation (Contd.)

SUMMARY FIGURE PROJECTED GLOBAL SALES FOR SMARPHONES, 2006‐2013 ($ MILLIONS)

• May 2009

BCC Research group report : “Global Market for Smart

120 140 160

Phones and PDAs” (USD 4850)

– 2008: Smart Phones market generated $58 7 billion

60 80 100 120

generated $58.7 billion – 2013: expected to increase to $153.3 billion

20 40

4 of 28

2006 2007 2008 2013

Objectives of the user identification system

1. Correct classification 2 Quick User identification 2. Quick User identification 3. Continuous monitoring 4. Resource efficient and light weight solutions solutions

5 of 28

Existing Methods Existing Methods

6 of 28

The solution???

K k D i

The solution???

• Keystroke Dynamics

40 45 40 45

• 20

25 30 35 40 R (%) 20 25 30 35 40 R (%) 5 10 15 20 FA 5 10 15 20 FR

7 of 28

Unacceptable method. Say goodbye to keystroke dynamics

The Classification Challenge The Classification Challenge

8 of 28

The Classification Challenge The Classification Challenge

• A problem of

A problem of Bio‐inspired classification classification

9 of 28

A Tri‐Mode System for User Identification

Tri‐Mode Tri Mode System Learning M d Detection M d Verification M d Mode Mode Mode

10 of 28

11 of 28

Learning Mode Learning Mode

• An optimizer fine tunes rule base and

database of a Fuzzy Classifier y

–Genetic Algorithm (GA)

• Darwinian Evolution
• Darwinian Evolution

–Particle Swarm Optimization (PSO)

• Feedback
• Feedback

– Hybrid of PSO and GA:

• Feedback controlled Darwinian Evolution

12 of 28

• Feedback controlled Darwinian Evolution

Detection Mode Detection Mode

• Fuzzy classifier trained and ready
• Continuous user monitoring

Continuous user monitoring

– We don’t know what will the user write

• Classification is done after e er 250 ke
• Classification is done after every 250 key

presses

– If the user is legitimate user, the system keeps on monitoring further

13 of 28

– If it raises as alarm, the system goes to verification mode

Verification Mode

• Activated when Detection Mode raises

Verification Mode

• Activated when Detection Mode raises

an alarm

• PIN based authentication

–match the typing behavior yp g

• we already know what is coming next

–Very accurate Very accurate

14 of 28

Verification Mode (The Maths) Verification Mode (The Maths)

15 of 28

Dataset

• A Key Logging application for Symbian

Dataset

• A Key Logging application for Symbian

based Nokia phones

• 25 users, 7 days

–From diverse backgrounds g –Includes students researchers professors people students, researchers, professors, people from corporate world, senior citizens businessmen engineers etc

16 of 28

citizens, businessmen, engineers etc

Visual Representation of features p

17 of 28

d Adjacent Non‐Adjacent Horizontal Digraph Adjacent Vertical Digraph Horizontal Digraph Vertical Digraph Non‐Adjacent Vertical Digraph

Visual Representation of features p

19 of 28

Experiments and Results

• Nature of Classification

p

– PSO‐GA‐Fuzzy scheme: Two class classification – Verification mode: Anomaly detection scheme

• For training we take 1 user as legitimate and 4

random users as imposters

• Testing done on remaining 20 users and the

legitimate user

• The user used as imposter in training is never

presented for testing

20 of 28

Accuracy Analysis y y

35 40 45 35 40 45 15 20 25 30 FAR (%) 15 20 25 30 FRR (%) 5 10 15 5 10 15

21 of 28

An improvement of 92.9% in FAR and 93.5% in FRR

Error Rate vs. Number of keys

18 20 u10 u14 u15 14 u10 u14 u15

y

12 14 16 18 10 12 6 8 10 12 6 8 FRR (%) FAR (%) 2 4 6 2 4 150 200 250 300 350 Number of Keypresses 150 200 250 300 350 Number of Keypresses

22 of 28

Training and Testing times

30 2.5

g g

20 25 e (secs) 1 5 2 e (secs) 10 15 aining Time 1 1.5 esting Time 5 Tra 0.5 Te

23 of 28

Limitations

• Identification delay of 250 keystrokes
• Identification delay of 250 keystrokes
• Accuracy sensitive to size of training data
• Not for QWERTY keyboard and Touch

Screen smart phones Screen smart phones

• Large training time
• Non‐Resilient to OS reinstallation

24 of 28

Overview of the Contribution

1 Identification of the problem domain as a bio inspired

• 1. Identification of the problem domain as a bio‐inspired

classification problem

• 2. A Keystroke‐based User Identification System for Smart

Phones with 93% improvement

• 3. Low runtime complexity ‐‐> Real world deployable
• 4. Dataset: will be released very soon

25 of 28

Acknowledgement g

• Information Communication Technology
• Information Communication Technology

Research and Development Fund (ICTR D F d) Mi i t f IT P ki t (ICTRnD Fund), Ministry of IT, Pakistan

www.ictrdf.org.pk

26 of 28

Screen shot of a desktop based d i ti d t derivative product

27 of 28