Identity verification and authentication (binding a human to an electronic transaction) have become strategic
business issues. How does a voice biometric system perform for a typical remote authentication business scenario, and what conclusions can we make about the
use of such a system?
(2007) Case Study: Phone-based Voice Biometrics for Remote Authentication
1. Case Study
Phone-based Voice Biometrics
for Remote Authentication
Stephen Elliot, Ph.D., Assoc Professor
Purdue University
&
Andy Rolfe, VP of Development,
Authentify Inc.
02/06/07 – ASEC-106
2. Objective
• Objective:
— Identity verification and authentication (binding a human
to an electronic transaction) have become strategic
business issues. How does a voice biometric system
perform for a typical remote authentication business
scenario, and what conclusions can we make about the
use of such a system?
4. Overview
• Briefly giving you an overview of:
— Biometric use in security systems
— The authentication best practices used
— The test methods
— Sample data
What we are NOT covering in presentation:
— Voice biometric or signal processing technology (FFT, HMM, etc.)
— Making any statement about the applicability of the technology for
your situation
19. Biometrics in Security
• Biometrics primer:
— Biometrics are by their nature statistically based
— Biometrics should not be the sole authenticator
— Backup methods for those that cannot (somehow impaired)
— Still have “first time” (registration) challenge
— Quality of implementation critical
• privacy,
• legal issues
• Multi-modal UI not easy
20. Voice Biometrics
• Why voice?
— Familiar paradigm; Very user acceptable; “business like”
— Multi-factor authentication in one session
— Real-time, undeniable contact for remote authentication
— Highly auditable
— Out of band trusted network
— Both physiological and behavioral
— Variable, dynamic samples
— No hardware deployment or training
21. Ease of Use & Intrusiveness (previous study)
70.00% 100.00%
60.00%
50.00% 80.00%
40.00% 60.00%
30.00%
20.00% 40.00%
10.00%
20.00%
0.00%
Not at all 4 3 2 Very 0.00%
Intrusive Intrusive Very Difficult Difficult Neutral Easy Very Easy
“I very much like the idea of voice
identification. This process
surpasses any other method of
protecting my identity and SSN that
I have seen. BRAVO!!
JoAnn W., Financial Advisory Firm
22. Security Best Practices
• Policies define process requirements
— Policy will (should) reflect risk profile
— Policy must account for risk for each factor of authentication
— Policy will define which factors will (should) be used & when
• Collect and use as many factors as possible
— Allows layering and substitution of factors depending on risks
• Fraudster may know everything about you,
but does not mean they can answer your telephone
23. Purdue Study
• Why study?
— No live system studies available
— Implementation specific
— Excellent resource nearby (Purdue University Biometrics Lab)
— Baseline for future studies
• biometric aging,
• technology changes,
• etc.
24. Biometric Comparisons
International Biometric Product Testing Initiative (May – Dec 2000) by National Physical Laboratory, England
[ sponsored by the Communications Electronics Security Group (CESG) ]
25. System used for Study
• This biometric study utilized a commercially available, remote,
service oriented security system.
• This system is actively being used by many corporations for mainly
Internet commerce and financial applications at a rate of
approximately 1.5M transactions per month.
• The test application was run using this active service environment
to best test "real life" performance of the technology.
• Test system implementation:
— SOA
— 2 step application
• Registration
• Verification
— Purdue University lab environment
26. Service Architecture
Engage the user, their computer and their telephone in a
synchronized exchange for a strong out-of-band authentication…
Users’ Web Session
Internet
Web
Servers Applet End
User
Corporate
Web Site
Bind the
https XML
Web session
the computer,
the phone and
the Person
Authentify
PBX
Service Ctr.
Public Switched Telephone Network 555-333-2399
( PSTN )
27. Roles & Responsibilities
• Authentify responsibilities:
— Design and implementation of enrollment & verification voice applications
— Operation of the commercial service center in Chicago
• Joint responsibilities
— Development of the test plan
— Data collection and reporting
— Data analysis and reports
• Purdue biometric lab responsibilities:
— Recruitment and instruction of test subjects
— Acquisition, operation and maintenance of equipment used by test subjects
— Provide assistance to ensure proper testing procedures
28. Biometrics Lab
• The Biometrics Lab at Purdue
is designed for research,
teaching, and testing
• Testing evaluation was
approved by the Institutional
Review Board at Purdue
University
• This research is typical of the
lab’s partnership with
company’s focusing on “applied
research”
• The lab is part of CERIAS
29. Test Protocol
• Data was collected at the Purdue University Biometrics Standards,
Performance, and Assurance Laboratory, in West Lafayette,
Indiana.
• The experimental area consisted of a room with minimal ambient
noise.
— Noise that was present was predominantly voices of other people, as
the room was utilized for other purposes during the experiment.
— Since more than one individual could do the study at the same time
and other individuals could be talking, noise conditions were
collected during the study.
30. Phones & Network Providers
• The land-based phone was a Vodavi • The Skype VoIP system used a
Starplus single line telephone. Linksys CIT200 Skype phone
— Land line provided by the university
• Cell phone services used:
• The Vonage VoIP system utilized a — T-Mobile
Linksys phone adapter and Uniden — Virgin Mobile
900 MHz cordless phone.
— Boost Mobile
— Network utilized was provided by the
university — Tracphone
— Network Speed 8,600 Kb/s upload / — Simple Freedom Wireless
86,000 Kb/s download
31. Data Capture
• The biometric system consisted of:
— Test subject web site where the sessions are initiated and the survey
results are captured
— Data capture enhancements to session processing
— Post processing of voice samples for more thorough test matrix
coverage
• Used combined speech recognition and speaker verification
• Used text prompted verification method (dynamic version of text
dependent verification)
• Did not use adaptation; did not test identification
32. Test Data
• Tests were automated to enable repeatable measurement of enrollment
and verification rates, and to capture the following data:
— Subject Identifier
— Trial Code (predetermined)
— Telephone Number
— Telephony Type (Landline, mobile, VoIP)
— Telephone Manufacturer & Model
— Telephone Location (address)
— Signal Strength (mobile phone only)
— Background Noise (Low | Med | High)
— Background Noise Type (Music | Speech | Noise)
— Subject’s Voice Health (Normal | Hoarse | Very Hoarse)
33. Data Analysis
• Data collection occurred in a indoor office environment
— Conversational background noise
• The test sessions captured all data utilized, so no preexisting
sample data was used.
• Enrollment templates and verification samples were compared both
in real-time and off-line after all test data had been collected.
• The combination of real-time sample capture and off-line
comparison helps generate a wider range of performance data.
34. Authentify-Purdue Study Results
Same Channel Performance -- Landline Verification vs. Landline Voiceprint
50.00%
45.00%
40.00%
35.00%
30.00%
Error Rate
25.00%
20.00%
15.00%
Land v Land
9.00%
10.00%
False Reject
5.00% 2.93% 3.61%
1.47% False Accept 0.49%
0.49%
0.00%
Low Med High
Security Level
35. Authentify-Purdue Study Results
Same Channel Performance -- Cell Verification vs Cell Voiceprint
50.00%
45.00%
40.00%
35.00%
30.00%
Error Rate
25.00%
20.00%
Cell v Cell
15.00%
12.87%
False Reject
10.00%
3.26% 2.63%
5.00%
False Accept
1.63% 1.08%
1.90%
0.00%
Low Med High
Security Level
36. Authentify-Purdue Study Results
Cross Channel Performance -- Cell Verification vs. Landline Voiceprint
50.00%
45.00%
40.00%
35.00% 37.43%
30.00%
False Reject
Error Rate
25.00%
Cell v Land
20.00%
15.00%
10.00% 11.90% 11.94%
5.00%
False Accept
0.00% 0.00% 0.00%
0.00%
Low Med High
Security Level
38. Conclusions
• Dynamic sampling is an effective method of supporting multi-factor
authentication in a single interaction
• Single voice biometric template capture OK for low to medium risk
applications when layered
• Best to use phone number or channel specific templates for
medium to high risk applications
• Use known phone number for verification to spawn new enrollment
session on secondary device (e.g. use existing landline print to
enroll on your new cell phone)
39. Conclusions
• We have got more work to do:
— Qualify batch analysis procedures
— Cell phone connection quality; how to compensate?
— VoIP is worst. Why?
— How much do behavioral characteristics play a role? Do subject
utterances change when they “know” they are acting as imposter?
— How well do biometric templates age? Use of adaptation?
— Can we leverage multiple verification engines to obtain a better
result?
— What role do accents play? Do they only affect reco’, or biometric
performance too?
40. Contact Information
Andrew Rolfe Stephen Elliott, Ph.D.
V.P. of Development & Operations Associate Professor & Director of
Biometric Standards, Performance,
and Assurance Laboratory
Phone: 773-243-0339 Phone: 765-494-1088
Email: andy.rolfe@authentify.com Email: elliott@purdue.edu
Authentify, Inc. Purdue University
8745 W. Higgins Road, Suite 240 401 N. Grant Street
Chicago, Illinois, 60631 West Lafayette, IN, 47906
www.authentify.com www.biotown.purdue.edu