1. Template Security in Biometric Systems
Yagiz Sutcu

2. Outline Introduction Biometrics and biometric systems
Template security and user privacy
Proposed solutions
Feature Transformation for ECC- based Biometric Template Protection
Syndrome framework and its requirements for binary data
Binarization of minutiae-based fingerprint templates
Protecting Biometric Templates with Secure Sketch
Secure sketch framework, issues and limitations
Quantization-based secure sketch
Randomization
Multi-factor setup
Conclusions, discussions and future directions

3. Biometrics
Biometrics is the science and technology of measuring and statistically analyzing biological data.
Adopted from: S. Prabhakar, S. Pankanti, and A. K. Jain, “Biometric Recognition: Security and Privacy Concerns”,
IEEE SECURITY & PRIVACY, 2003.
Universality - do all people have it ?
Distinctiveness: can people be distinguished based on an identifier ?
Permanence : how permanent is the identifier ?
Collectability : how well can the identifier be captured and quantified ?
Performance : speed and accuracy
Acceptability : willingness of the people to use
Circumvention : foolproof

4. Biometrics

5. Biometric Systems
S. Prabhakar, S. Pankanti, and A. K. Jain, “Biometric Recognition: Security and
Privacy Concerns”, IEEE SECURITY & PRIVACY, 2003.

7. Issues, Challenges and Objectives
Biometric authentication is attractive
Closely related to the identity
Cannot be forgotten
Not easy to forge
Have been successfully used for a long time
However …
Cannot be exactly reproduced (intra-variability, noise)
Once compromised cannot be revoked
Entropy may not be sufficient
Objectives
Authentication/verification without storing the original biometric
Robustness to noisy measurements
Best possible tradeoff between
Security (How many bits must an attacker guess?)
Accuracy/performance (What is the false reject rate?)

8. Proposed Solutions – Transformation-Based Approaches
Employ one-way transformation
E.g., quantization, thresholding, random projections
Properties
Non-invertible or hard-to-invert
Similarity preserving
Cancelable
Technique depends highly on the biometric considered
Security not easy to analyze
Ratha’01&’07, Savvides’04, Ang’05, Teoh’06, etc.

9. An Example: Cancelable Biometrics
Repeatable
Distortion
Match
Do not Match
Applied at the sensor level
Signal level
Feature level
Security
If compromised, a new distortion
Reusability
Different distortions for different databases
Performance
What about false accept and false reject rates?
Requires alignment

10. Proposed Solutions – Helper Data-Based Approaches
Generate user-specific helper data
E.g., syndrome, secure sketch
Helper data and generation method are public
General ECC-based framework that is applicable to many biometrics
Techniques may vary to optimize performance for different modalities
Security analysis based on information-theory is possible
Davida’98, Juels’99&’02, Dodis’04, Martinian’05, Draper’07, etc.

11. An example: “Fuzzy Commitment”
If the noisy biometric (X’) is close enough to the template (X), decoder successfully corrects the error.
The only information stored are  and hash(K)

12. An Example for Helper Data-based Biometric Template Protection: Syndrome Coding
… About securing biometric templates using error correction codes.
… Joint work

13. Syndrome Coding Framework
Encode
enrollment
biometric
Decode
w/ probe
biometric
Store syndrome S and hash of X X S
Syndrome
Encoding
Syndrome
Decoding
Biometric
Authentication
Authenticate
only if hash of
estimate matches
stored hash
Original
enrollment biometric Y
Fingerprint
Channel
Noisy biometric probe
S cannot be uncompressed by itself and is therefore secure
In combination with a noisy second reading Y the original X can be recovered using a Slepian-Wolf decoder
Compare hash of estimate with stored hash to permit access

14. System Implementation
Enrolment
Fingerprint
access
granted
Alignment and Minutiae
Extraction
Extract
binary feature
vectors
Syndrome
Encoding
yes no
access
denied
Syndrome
Database
yes
Alignment and Minutiae
Extraction
Extract
binary feature
vectors no
access
denied
Syndrome
Decoding
Probe
Fingerprint

15. Overview of Syndrome Encoding/Decoding
Security = number “missing” bits
= original bits – syndrome bits
Translates into number guesses
to identify original biometric w.h.p.
Robustness = false-rejection rate
Robustness to variations in biometric readings
achieved by syndrome decoding process
(syndrome + noisy biometric => original biometric)
Fewer syndrome bits = greater security, but less robustness

16. Example: Distributed Coding with Joint Decoding
Syndrome
Bits
Source X
Source Y 1 1 1 1 1 1

17. Example: Syndrome Decoding
Bits
Source X
Side Info Y ? 1 ? 1 ?
Use side info Y and
syndrome bits to
reconstruct X ? ? 1 ? ?

18. Example: Syndrome Decoding
Bits
Source X
Guess X 1
Flipping 1st bit from 0 to 1 satisfies 1st syndrome but
violates 2nd syndrome 1 1 1 1 1 1

19. Example: Syndrome Decoding
Bits
Source X
Guess X 1
Flipping 2nd bit from 1 to 0 satisfies syndrome bits 1 and 2 but violates 3rd syndrome bit 1 1 1 1

20. Example: Syndrome Decoding
Bits
Source X
Guess X 1 1 1
Flipping 3rd bit from 0 to 1 satisfies all syndrome bits and recovers X 1 1 1 1

21. Security of Syndrome Approach
list of biometrics
satisfying linear
constraints
F(X), set of linear functions
specified by code C 1 1 1 1
F (F(X)) -1 + 1 + …. +
secure biometric, S = evaluation
of functions (syndrome vector)
enrollment
biometric X

22. Previous Approach: Binary Grid Representation
Problems
Representation is sparse and difficult to model
Statistics of binary string are not well suited for existing codes
Poor performance
Solution: Pre-processing!
Pre-process fingerprint data to produce a binary string that is statistically compatible with existing codes
Syndrome coding on resulting binary string 1

23. Desired Properties of Feature Transformation. 1
Zeros and ones
equally likely
Individual bits
independent
User A
User B
Independent
bit strings
Reading 1
Reading 2
BSC-p 2 3 4
1. Bits extracted from biometrics should have statistical properties which make the biometric channel simple
2. List the properties and implications. Entropy = 1 bit, Pairwise entropy = 2 bits
3. Then the recipe for a secure biometrics scheme would be…

24. Extracting Robust Bits from BiometricsY 
N random cuboids in
minutiae map
# minutiae
Bit vector 6 7 9 1 +
Median
thresholds
Each cuboid contributes a 0 or 1 bit to the feature vector, if it contains
less or more minutia points than the median

25. Elimination of Overlapping Cuboids
Large overlap  similar bits  easy for attacker to guess
Large overlap results in low pair-wise bit-entropy, i.e., easier to guess pairs of bits
To prevent, reject cuboids which result in low pair-wise entropy with other cuboids.
Training process done only once for the complete database.
400 cuboids
Best 150 cuboids

26. User-Specific Cuboids
Different users have different set of cuboids
Requirements: For each user, choose the cuboids which
Are most reliable
Result in equal number of zeros and ones
Have smallest possible pairwise correlation
What is a “reliable” cuboid?
# minutiae points far away from the median
0 or 1 bit is likely to remain unchanged over repeated
noisy measurements of fingerprint.

27. “Reliable” Cuboids 1 UNRELIABLE RELIABLE
8 minutiae
3 minutiae
Measurement 1
Measurement 1
5 minutiae
9 minutiae
Measurement 2
Measurement 2
For robust bit extraction, we would prefer the cuboids to be “reliable”
Consider a cuboid, median number of minutia = 4
Run through the example
UNRELIABLE
RELIABLE 1
E.g., median = 4
Each user has a different set of reliable cuboids !

28. User-Specific Reliable Cuboids
0-List
1-List …
Sort by reliability
N fair coin flips to
choose cuboids
from top of each list 1 2 3 4 5
Unordered list
How to construct a feature vector using the most reliable cuboids?
List of 0- and 1-cuboids
Farther the number of minutia points is from the median, the more reliable the cuboid is. Sort.
Chose cuboids from the top of the list at random
Verify 4 properties:
Fair coin = nearly equal 0s and 1s
Independent coin tosses = independent bits
Different users have different reliable cuboids = independent bitstreams across users
Reliable cuboids = intra-user bit flips have small probability

29. Distribution of Zeros and Ones
Proprietary database of 1035 users, 15 pre-aligned samples per user
Desired 75 ones
Desired 75 ones 20 40 60 80
100
120
140 50
150
200
250
Number of 1's in the transformed feature vectors
Number of feature vectors 30 40 50 60 70 80 90
100
110
120
150
200
250
300
350
400
450
Number of 1's in the transformed feature vectors
Number of feature vectors
Clustered around 75. i.e., half the number of 1’s because we chose randomly from the two ordered lists of 0 and 1 Cuboids.
150 Common Cuboids
150 User-Specific Cuboids

30. Intra-User and Inter-User Separation
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
0.02
0.04
0.06
0.08
0.12
0.14
0.16
0.18
Distribution of the NHD
Normalized Hamming Distance (NHD)
intra-user variation
inter-user or
attacker variation
0.35
inter-user variation
0.3
0.25
Distribution of the NHD
0.2
intra-user variation
attacker steals
user’s cuboids
0.15
NHD is used as a distance criterion
Overlap between intra-user and inter-user distributions
For user-specific case, overlap is dramatically reduced because every user has different set of reliable cuboids
Green plot => what if attacker gets access to the reliable user-specific questions (worst case scenario). How much does that compromise security
0.1
0.05
0.2
0.4
0.6
0.8 1
Normalized Hamming Distance (NHD)
150 Common Cuboids
150 User-Specific Cuboids

31. Equal Error Rates
User-specific cuboids provide lower equal error rate even if the attacker knows everybody’s cuboids.
0.05
0.1
0.15
0.2
0.25
Intra-user NHD
Inter-user NHD
0.02
0.04
0.06
0.08
0.12
0.14
0.16
0.18
Inter-user NHD or attacker NHD
inter-user scenario
attack scenario
0.027
≈ 0
ROC curves, EER
If the attacker knows the questions, the EER increases. But it is still less than the EER for common cuboids.
Common Cuboids
User-Specific Cuboids

32. Syndrome Coding Results

33. Conclusions/Discussions
Random cuboids enable robust bit extraction with desired
properties
User-specific (reliable) features require more computation and storage, but give better separation between intra-user and inter-user feature vectors and provide higher security than common feature vectors
However…
Fast method for eliminating correlated bit-pairs from user-specific cuboids
Extending feature transformation to use ridge data which is provided along with minutiae map
Observing effect of alignment inconsistencies on overall performance

34. Protecting Biometric Templates with Secure Sketch: Theory and Practice
… About securing biometric templates using error correction codes.
… Joint work

35. Secure Sketch ENCODER DECODER Enrollment Verification Noise P Sketch
randomness
Generate
password/key
Sketch should not reveal too much information about the original biometric

36. Secure Sketch Entropy-loss (min-entropy of X)
(average min-entropy of X given P)

37. Security of secure sketch is defined in terms of entropy loss, L
Suppose original biometric data X have min-entropy H(X)
The strength of the key we can extract from X given sketch P is at least H(X) – L
L can be easily bounded by the size of the sketch P
L is an upper bound of information leakage for all distributions of X

38. However… Known secure sketch schemes have limitations
Only deal with discrete data
But real world biometric features are mostly in continuous domain
One general solution
Quantize/discretize the data and apply known schemes in the quantized domain

39. Quantization-based Secure Sketch

40. A Simple Example
X is original data, 0 < X < 1 and under noise, X can be shifted by at most 0.1 1 1

41. Problems Remain... For different quantization methods
Min-entropy of quantized data could be different
Entropy loss could also be different
How to define the security?
Using entropy loss in the quantized domain?
Could be misleading

42. Different Quantizers Using scalar quantizer: Which one is better?
Case 1:
step size 0.1
entropy loss = log 3
Case 2:
step size 0.01
entropy loss = log 11
Which one is better?
It depends on distribution of X
If X is uniform:
Case 1: H(X) = log(100), H(X) – L = log (100/3) = 5.06
Case 2: H(X) = log(1000), H(X) – L = log(1000/11) = 6.51
Case 2 yields a stronger key
However, there exists a distribution of X such that for both case 1 and 2:
H(X) is the same
L is the actual information leakage
Case 1 yields a stronger key

43. How to Compare?
Or, how to define the ``optimal'' quantization for all distributions of X?
It is difficult
Might be impossible
Instead, we propose to look at relative entropy loss, in addition to entropy loss
Essentially, we ask questions differently:
Given a family of quantizers (say, scalar quantizers with different quantization steps), for any one of them (say, Q), how many bits more that we could extract from X if another quantizer Q' was used?
How to bound the number of additional bits that can be extracted for any Q' (compared with Q)?
If we can bound it by B, then the ``best'' quantizer in the family cannot be better than Q by more than B bits

44. Main Results
For any well-formed quantizer family, we can always bound the relative entropy loss
well-formed: no quantizer in the family loses too much information (say, having too large a quantization step)
The safest way to quantize data is to use a quantization step same as the error rate
safest: relative entropy loss is the smallest
This result is consistent with intuition
useful to guide practical designs

45. However… Known secure sketch schemes have limitations
Only deal with discrete data
But real world biometric features are mostly in continuous domain
One general solution
Quantize/discretize the data and apply known schemes in the quantized domain
Measure security using entropy loss in the quantized domain
For different quantization methods
Min-entropy/entropy could be different
How to define the security?
Using min-entropy alone could be misleading
Improve performance?
How about cancelability/reusability?
Better feature selection?

46. Randomized Quantization-based Secure Sketch
Enrollment
Verification X XR
Q(XR)
ENCODER
randomization
quantization Y YR
Q(YR)
DECODER PX
sketch
template
noisy
biometric

47. Results
ORL Face database - 40 different individual and 10 samples per individual
7 for training and 3 for testing
PCA features (eigenfaces) considered
Range-based similarity measure
User-specific random (uniform) matrices

48. Results
randomization
quantization

49. Results
Min-entropy
Average sketch-size
Left-over entropy
PCA-based selection
59.91
40.35
19.56
based
selection
61.95
43.46
18.49
132.52
98.88
33.64
139.95
106.55
33.40
n=20
n=50

50. Conclusions/Discussions
Randomization
Improve performance
Cancelability/reusability
Feature selection
Similar security with better average sketch-size estimation
However…
How to measure biometric information?
Entropy estimation
Matching algorithm
How to define/find the “Optimal” quantization?
Given the input distribution
Given practical constraints (size of sketch and/or templates)
Different quantization strategies

51. Overview of Future Directions

52. Threats/Attack Vectors
Chris Roberts, “Biometric attack vectors and defences”, Computers & Security 26(1): (2007)

53. Secure Biometric Systems: A Blend of Multiple Disciplines
Feature extraction/matching with robustness to noise and other variations;
Finding new/better features
Pattern Recognition
Signal Processing
Cryptography
Recovery of original data
from noisy or corrupt data;
Better transformations
Protect/scramble biometric
data against malicious attacks;
Analysis of security levels
offered by various techniques.
Secure
Biometric
Systems

54. Open Issues/Research Opportunities
Improving robustness vs. security trade-off
Reliable measurements of biometric information
Inherent entropy of biometrics
Connection between template security and system security
Remote vs. local
Two-party vs. multi-party
Multi-biometrics
Multi-factor
User privacy
Standardization
Terminology
Format

55. Thanks